A multi-modal data security analysis technology based on a deep fuzzy wavelet learning model

Abstract

The application provides a multi-modal data security analysis technology based on a deep fuzzy wavelet learning model. The deep fuzzy wavelet learning model is widely used in the field of security analysis due to its self-adaptive ability in complex data environment. The method improves the deep fuzzy wavelet learning framework and applies it to multi-modal data security analysis. The model first uses the method combining wavelet transform and fuzzy clustering to perform feature decomposition and noise reduction processing on multi-modal data (text, speech, video, image), extracts key semantic, timing and spatial features, and provides high-dimensional information representation for subsequent security analysis. Secondly, a deep fuzzy wavelet neural network is constructed, the extracted features are input into the adaptive fuzzy decision layer, and the wavelet learning model is combined to capture the spatio-temporal correlation features, improve the perception ability of the model to complex attack patterns. Finally, the security analysis result is strengthened by using the space search optimization algorithm, so as to effectively detect malicious samples, identify backdoor attacks, and improve the defense ability against attacks, and finally realize accurate multi-modal data security evaluation.

Description

Technical Field

[0001] This invention relates to the field of data security analysis, specifically to a method for deep fuzzy wavelet learning. Background Technology

[0002] Multimodal data security analysis is a crucial application of artificial intelligence in cybersecurity. It identifies potential security threats by analyzing and processing data from different modalities, thereby enhancing system defense capabilities and reducing security risks. In recent years, with the rapid development of technologies such as deep learning, computer vision, and natural language processing, multimodal data security analysis has made significant progress. However, real-world cybersecurity threats exhibit high complexity and stealth. When faced with new attack methods or changes in the data environment, existing detection models may struggle to accurately identify anomalous patterns because new data may contain attack features or malicious behavior patterns that the model has not encountered before. This necessitates continuous model updates and adjustments, leading to increased computational resource consumption. Deep fuzzy wavelet learning models, due to their powerful feature extraction and adaptive learning capabilities, have broad application prospects in multimodal data analysis. However, traditional methods still face challenges in handling the fusion of different modalities. One major problem is the heterogeneity of feature representations between different modalities, which limits the effectiveness of information sharing and joint analysis, thus affecting the model's ability to identify complex attack behaviors. Summary of the Invention

[0003] The purpose of this invention is to overcome the above-mentioned technical shortcomings and deficiencies by providing a multimodal data security analysis technology based on a deep fuzzy wavelet learning model.

[0004] The technical steps of this invention are as follows:

[0005] 1. A multimodal data security analysis technique based on a deep fuzzy wavelet learning model, comprising the following steps:

[0006] Step 1: Obtain multimodal security analysis datasets. The experimental data uses the MIDAS and AVID datasets. The MIDAS dataset contains text logs, network traffic, and Industrial Control System (ICS) data, which can be used to detect malicious attacks; the AVID dataset contains adversarial examples for images and videos, suitable for adversarial attack detection. In the experiments, preprocessing and feature extraction were performed on different modalities of data. Text data was segmented, and semantic features were extracted using TF-IDF and BERT embedding methods. Image data was decomposed using multi-scale wavelet transform (DWT) to extract low-frequency and high-frequency features, and depth features were extracted using a ResNet-50 convolutional neural network (CNN). Mel-frequency coefficients (MFCC) were calculated for audio data to obtain time-frequency features. Keyframes were extracted from video data, and temporal features were extracted using a 3D convolutional network (C3D).

[0007] Step 2: Construct a deep fuzzy wavelet learning model based on the fusion of wavelet neural network (WNN) and fuzzy inference system (FIS). Information gain is used to select key features, avoiding the influence of invalid features on network training. In the model, wavelet transform is used to enhance data resolution, fuzzy membership functions are used to model uncertainty, and adaptive optimization algorithms are used to optimize network parameters, improving the accuracy of multimodal data security analysis.

[0008] Step 3: Optimize the model using a spatial search optimization algorithm to reduce fuzzy overlap between different categories and improve classification accuracy. Specifically, the initial network is used as a baseline, and global search, local search, and contrastive search strategies are employed to iteratively optimize the population, thereby improving the model's generalization ability and robustness, and enhancing its ability to detect malicious samples.

[0009] Step 4: Output multimodal data security analysis results. Input the multimodal data to be detected, use the optimized deep fuzzy wavelet learning model for classification and anomaly detection, identify potential security threats, including data poisoning, adversarial attacks, etc., and output the analysis results.

[0010] The specific steps for step 1 are as follows:

[0011] Step 1.1: The datasets used in this experiment are the MIDAS dataset and the AVID dataset. The MIDAS dataset contains 12,000 ICS logs, 8,500 network traffic data entries, and 6,300 text logs, covering normal operational data and various attack samples. The AVID dataset contains 5,000 adversarial sample images and 3,200 adversarial sample videos, covering security threats such as noise attacks, steganography attacks, and data poisoning.

[0012] Step 1.2 Data Preprocessing: For different modalities of data, denoising, format conversion, and standardization are performed respectively: For text data, stop words are removed, stemming is performed, and BERT word segmentation is used for sub-word decomposition; For image data, the resolution is unified to 224×224, normalization is performed, and the image is converted to Lab color space to enhance edge features; For audio data, short-time Fourier transform (STFT) is used for time-frequency analysis, and silent segments are removed; For video data, key frames are extracted, and optical flow is used to calculate motion information.

[0013] Step 1.3: Text data feature extraction. TF-IDF and BERT embedding methods are used to obtain text semantic features.

[0014] TF-IDF calculation characteristics:

[0015]

[0016] Among them, f t Let n be the number of times word t appears in document d, and N be the total number of documents. t t represents the number of documents containing t.

[0017] Step 1.4: Image data feature extraction, using wavelet transform (DWT) and ResNet-50 for feature extraction.

[0018] Wavelet transform characteristics:

[0019] Perform multi-scale wavelet decomposition on the image to extract low-frequency components (LL) and high-frequency components (LH, HL, HH):

[0020]

[0021] Where s[m,n] is the input image, These are wavelet basis functions.

[0022] Step 1.5: Audio data feature extraction, calculation of Mel spectral coefficients (MFCC) and spectral centroid.

[0023] MFCC calculation:

[0024]

[0025] Among them, X m is the spectrum after Mel filtering, and k is the coefficient index.

[0026] Spectral centroid calculation:

[0027]

[0028] Where S(f) is the power spectrum at frequency f.

[0029] Step 1.6: Video data feature extraction. Optical flow method is used to calculate inter-frame motion information and extract spatiotemporal attention features.

[0030] Features of optical flow calculation:

[0031]

[0032] Where u and v represent the components of optical flow in the x and y directions, respectively, and I is the image pixel intensity.

[0033] Step 1.7: Feature Normalization and Dimensionality Reduction

[0034] Normalization: The Z-score standardization method is used to make the feature distribution mean 0 and variance 1.

[0035]

[0036] Where μ is the mean and σ is the standard deviation.

[0037] Principal component analysis is then used to reduce feature dimensions and improve computational efficiency.

[0038] The specific steps for step 2 are as follows:

[0039] Step 2.1: Feature selection based on information gain

[0040] The features extracted from the MIDAS and AVID datasets are filtered, the information gain of each feature is calculated, and redundant or invalid features are removed.

[0041] The formula for calculating information gain is as follows:

[0042] IG(T,) = H(T) - H(T|X)

[0043] Where IG(T,X) represents the information gain of feature X on target variable T, H(T) is the entropy of target variable, and H(T|X) is the conditional entropy of T given feature X.

[0044] Step 2.2: Constructing a fuzzy wavelet neural network

[0045] The feature space is partitioned using fuzzy C-means (FCM) clustering, and fuzzy membership degrees are calculated:

[0046]

[0047] Among them, u ij For sample x j Membership degree of cluster i, v i Let m be the cluster center and m be the fuzziness index.

[0048] Then, wavelet transform is used for data feature enhancement, and the obtained fuzzy features are decomposed into multi-scale features:

[0049]

[0050] Among them, W ψ (a,b) are wavelet coefficients. Let a and b be wavelet basis functions, and a and b represent the scaling and translation parameters, respectively.

[0051] Step 2.3: Optimize the fuzzy rules and model parameters using an adaptive optimization algorithm.

[0052] The adaptive optimization algorithm is used to optimize fuzzy rules and network parameters to improve the model's learning efficiency and adaptability. This includes the following steps:

[0053] Adaptive rule adjustment:

[0054] A data-driven approach is employed to dynamically adjust the fuzzy rules in the fuzzy wavelet neural network to adapt to changes in the distribution of different categories. The importance is determined by I... r express:

[0055]

[0056] Among them, u rj For sample X j The membership degree of rule r, IG(T,X) j ) represents the information gain of this feature.

[0057] Next, the Adam optimization algorithm is used to optimize the adaptive learning rate:

[0058] Adam optimization algorithm:

[0059]

[0060] Where, θ t Here are the current model parameters, L is the loss function, α is the initial learning rate, and m... t and v t The first and second moment estimates are given respectively, and β1 and β2 are the exponential decay rates.

[0061] Next, the learning rate is dynamically adjusted based on changes in loss to enhance adaptability to multimodal data and perform effective feature selection. The adjustment formula is as follows:

[0062]

[0063] Where, η t λ represents the current learning rate, and λ is the decay coefficient.

[0064] The specific steps for step 3 are as follows:

[0065] Step 3.1: Using the initial model trained in Step 2 as a baseline, construct an optimization population and initialize multiple possible parameter combinations to ensure that the optimization search range covers a sufficient feature space.

[0066] Assume the model parameter vector is:

[0067] Θ = {θ1, θ2, ..., θ n}

[0068] Where, θ i The i-th parameter of the model is used to initialize the population {Θ}. 1 ,Θ 2 ,...,Θ M} serves as the starting point for optimization, where M is the population size.

[0069] Step 3.2: Use a global search strategy to randomly sample in the high-dimensional feature space, with the goal of minimizing the classification error loss function.

[0070]

[0071] Where N is the training sample, x i For training data, y i For the real label f(x) i ;Θ) represents the model's predicted output. The cross-entropy loss function is optimized using a gradient estimation-based search strategy.

[0072]

[0073] Where η is the learning rate. This represents the gradient of the loss function with respect to the parameters.

[0074] Step 3.3: Based on the excellent individuals obtained from the global search, a local search is introduced. Gradient information or heuristic methods are used to fine-tune the model parameters to reduce randomness in the optimization process and improve classification stability. Dynamic gradient descent is used to accelerate convergence.

[0075]

[0076] Θ t+1 =Θ t +v t+1

[0077] β is the momentum factor, which makes the parameter updates affected by the gradient of the previous time step, thus smoothing the optimization path.

[0078] Step 3.4: Introduce an adversarial search strategy, construct adversarial features or data augmentation strategies, conduct robustness tests on the model, and optimize the adversarial loss to improve the model's ability to detect abnormal and attack samples.

[0079] To improve the model's ability to detect adversarial attacks, an adversarial perturbation δ is added to generate adversarial examples:

[0080]

[0081] Where ∈ represents the perturbation amplitude, sign(·) represents the gradient direction, and the generated x adv This can be used to evaluate the adversarial robustness of the model. The optimization objective becomes:

[0082]

[0083] The max operation finds the most difficult adversarial examples to detect, while the min operation optimizes the model's robustness to them.

[0084] Step 3.5: Evaluate the performance of the optimized model, comparing classification accuracy, robustness metrics, and computational cost before and after optimization. Finally, select the optimal parameter combination to complete the optimization process. Evaluation metrics are as follows:

[0085] Classification accuracy:

[0086]

[0087] Where 1(·) is an indicator function, These are predicted values.

[0088] Adversarial example detection success rate:

[0089]

[0090] in, For adversarial sample x adv The prediction results, ASR, reflect the robustness of the model.

[0091] Ultimately, the optimal parameter Θ that achieves the highest classification accuracy and lowest ASR was selected as the final model parameter.

[0092] Step 4 outputs the security analysis results, which involves detecting the input multimodal data and outputting the data content security analysis results.

[0093] Advantages and benefits of the present invention

[0094] This invention proposes a multimodal data security analysis technique based on a deep fuzzy wavelet learning model. This technique can effectively fuse data from different modalities, improving the accuracy of security threat detection. First, this invention employs a deep fuzzy wavelet learning model to extract and fuse features from multimodal data such as text, images, audio, and video, thereby enhancing the ability to identify complex security threats. Second, during model training, an information gain method is used to select key features, avoiding the influence of invalid features on network training and improving the model's learning efficiency and detection accuracy. Furthermore, this invention utilizes a spatial search optimization algorithm to optimize the model, reducing fuzzy overlap between different categories and improving classification accuracy and generalization ability. Experimental results on the MIDAS and AVID datasets demonstrate that this invention exhibits superior detection performance in multimodal data security analysis tasks. Attached Figure Description

[0095] Figure 1 This is a flowchart illustrating the overall implementation of a deep fuzzy wavelet learning model for multimodal data security.

[0096] Figure 2 This is the overall flowchart of the deep fuzzy wavelet learning model; Detailed Implementation

[0097] The present invention will now be described in detail with reference to the accompanying drawings and examples.

[0098] Step 1: Obtain the multimodal security analysis dataset. The experimental data uses the MIDAS and AVID datasets. In the experiment, preprocessing and feature extraction were performed on different modalities of data. Text data underwent word segmentation, and semantic features were extracted using TF-IDF and BERT embedding methods. Image data was decomposed using multi-scale wavelet transform (DWT) to extract low-frequency and high-frequency features, and depth features were extracted using a ResNet-50 convolutional neural network (CNN). Mel-frequency coefficients (MFCC) were calculated for audio data to obtain time-frequency features. Keyframes were extracted from video data, and temporal features were extracted using a 3D convolutional network (C3D). The specific steps are as follows:

[0099] Step 1.1: Randomly sample from the MIDAS and AVID datasets. The MIDAS dataset contains 12,000 ICS logs, 8,500 network traffic data points, and 6,300 text logs; the AVID dataset contains 5,000 adversarial example images and 3,200 adversarial example videos. Divide the dataset into training and testing sets, with 70% used for training and 30% for testing.

[0100] Step 1.2: Preprocess the text data, including stop word removal, stemming, and word decomposition using the BERT word segmenter. The final result is converted into TF-IDF feature vectors (5000 dimensions) and BERT embedding vectors (768 dimensions). The training set is 9800×(5000+768), and the test set is 4200×(5000+768).

[0101] Step 1.3: Preprocess the image data. First, resize the image to 224×224 and convert the color space to Lab color space to enhance edge features. Then, apply Multi-Scale Wavelet Transform (DWT) to decompose the image, extracting low-frequency and high-frequency components, and combine this with ResNet-50 for feature extraction. The final feature vector dimension is 2048, the training set dimension is 3500×2048, and the test set dimension is 1500×2048.

[0102] Step 1.4: Preprocess the audio data, perform time-frequency analysis using short-time Fourier transform (STFT), remove silent segments, then calculate the Mel spectral coefficients (MFCC) to obtain 40-dimensional features, and combine 5 additional features such as the spectral centroid, the final feature dimension is 45, the training set is 2240×45, and the test set is 960×45.

[0103] Step 1.5: Process the video data. First, optical flow is used to calculate inter-frame motion information, and 16 keyframes are extracted from each video segment. Each frame is 112×112×3 in size. Then, a 3D convolutional network (C3D) is used to extract temporal features. The final feature vector dimension is 4096, the training set is 2240×4096, and the test set is 960×4096.

[0104] Step 1.6: Perform Z-score normalization on all features to achieve a mean of 0 and a variance of 1, and then use PCA for dimensionality reduction to ultimately unify the feature dimensions to 512 dimensions, thereby improving computational efficiency. The training set has a dimension of 9800×512, and the test set has a dimension of 4200×512.

[0105] Step 2: Construct a deep fuzzy wavelet learning model based on the fusion of wavelet neural network (WNN) and fuzzy inference system (FIS). Information gain is used to select key features, avoiding invalid features from affecting network training. In the model, wavelet transform is used to enhance data resolution, fuzzy membership functions are used to model uncertainty, and adaptive optimization algorithms are used to optimize network parameters, improving the accuracy of multimodal data security analysis. Specific steps are as follows:

[0106] Step 2.1: Calculate the information gain of all features and select the 128 features with the highest contribution to reduce the impact of redundant features on the training process.

[0107] Step 2.2: The selected 128-dimensional features are divided into 5 clusters using the Fuzzy C-means (FCM) clustering method, and fuzzy membership degrees are calculated. Subsequently, wavelet transform is applied to perform multi-scale decomposition of the features, extracting high-frequency and low-frequency features, which are then combined to form a new feature representation. The training set dimension is 9800×256, and the test set dimension is 4200×256.

[0108] Step 2.3: Train the WNN model using the Adam optimization algorithm. The initial learning rate is set to 0.001, decreasing by a factor of 0.5 every 10 rounds, with a maximum of 100 training rounds and a batch size of 128. The loss function is the mean squared error (MSE).

[0109] Step 3: Optimize the model using a spatial search optimization algorithm to reduce fuzzy overlap between different categories and improve classification accuracy. Specifically, the initial network is used as a baseline, and global search, local search, and contrastive search strategies are employed to iteratively optimize the population, thereby improving the model's generalization ability and robustness, and enhancing its ability to detect malicious samples. The specific steps are as follows:

[0110] Step 3.1: Initialize 50 different combinations of hyperparameters, including learning rate (0.001~0.01), number of fuzzy rules (5~10), wavelet basis functions (Haar, Daubechies, Symlets), and construct a population for search optimization.

[0111] Step 3.2: Use a global search strategy to perform preliminary parameter screening and calculate the cross-entropy loss. The optimal parameter combination is: learning rate 0.005, wavelet basis function Daubechies-4, number of fuzzy rules 7, and the classification accuracy reaches 92.4%.

[0112] Step 3.3: Based on the global search, a local search strategy is applied, and a momentum factor of 0.9 is used for gradient update to further optimize the classification model, ultimately improving the classification accuracy to 93.8%.

[0113] Step 3.4: Introduce an adversarial search strategy to generate adversarial examples. Use the FGSM method to add perturbation to the input data (perturbation amplitude ε = 0.01), calculate the model's detection success rate on adversarial examples, optimize the adversarial loss, and finally reduce the ASR (adversarial attack success rate) to 6.2%.

[0114] Step 3.5: Evaluate the final model. The classification accuracy is 94.2% and the adversarial example detection success rate is 93.1%. Finally, determine the optimal hyperparameter combination and complete the model training.

[0115] Step 4 outputs the security analysis results, which involves detecting the input multimodal data and outputting the data content security analysis results.

[0116] The deep fuzzy wavelet learning model proposed in this invention was tested eight times using the MIDAS and AVID datasets, and the average results were analyzed. The main evaluation metrics were classification accuracy (%), robustness index (RI), and adversarial attack error rate (AER, %), with FMNN (fuzzy neural network) as the baseline. The experimental results are shown in Table 1.

[0117] Table 1 Comparison of Data Results

[0118]

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the invention should be included within the scope of the claims.

Claims

1. A multimodal data security analysis method based on a deep fuzzy wavelet learning model, comprising the following steps: Step 1: Obtain multimodal security analysis datasets. The experimental data uses the MIDAS and AVID datasets. The MIDAS dataset contains text logs, network traffic, and Industrial Control System (ICS) data, which can be used to detect malicious attacks. The AVID dataset contains adversarial examples for images and videos, suitable for adversarial attack detection. In the experiments, different modal data are preprocessed and feature extracted. Text data is segmented, and semantic features are extracted using TF-IDF and BERT embedding methods. Image data is decomposed using multi-scale wavelet transform (DWT) to extract low-frequency and high-frequency features, and depth features are extracted using a ResNet-50 convolutional neural network (CNN). Mel-frequency coefficients (MFCC) are calculated for audio data to obtain time-frequency features. Keyframes are extracted from video data, and temporal features are extracted using a 3D convolutional network (C3D). Step 2: Construct a deep fuzzy wavelet learning model based on the fusion of wavelet neural network (WNN) and fuzzy inference system (FIS); use the information gain method to select key features and avoid invalid features from affecting network training; In the model, wavelet transform is used to enhance data resolution, fuzzy membership function is used to model uncertainty, and network parameters are optimized through adaptive optimization algorithm to improve the accuracy of multimodal data security analysis; Step 3: Optimize the model using a spatial search optimization algorithm to reduce fuzzy overlap between different categories and improve classification accuracy. In the specific optimization process, the initial network is used as a baseline, and global search, local search and opposition search strategies are used to optimize and iterate the population, thereby improving the model's generalization ability and robustness, and enhancing the ability to detect malicious samples. Step 4: Output multimodal data security analysis results. Input the multimodal data to be detected, use the optimized deep fuzzy wavelet learning model for classification and anomaly detection, identify potential security threats, including data poisoning, adversarial attacks, etc., and output the analysis results.

2. The multimodal data security analysis method based on a deep fuzzy wavelet learning model according to claim 1, characterized in that: The specific steps for acquiring multimodal data and extracting features in step 1 are as follows: Step 1.1: The datasets used in this experiment are the MIDAS dataset and the AVID dataset. The MIDAS dataset contains 12,000 ICS logs, 8,500 network traffic data, and 6,300 text logs, covering normal operation data and various attack samples. The AVID dataset contains 5,000 adversarial sample images and 3,200 adversarial sample videos, involving security threats such as adversarial noise, steganography attacks, and data poisoning. Step 1.2 Data Preprocessing: For different modalities of data, denoising, format conversion and standardization are performed respectively: For text data, stop words are removed, stemming is performed, and BERT word segmentation is used for sub-word decomposition; For image data, the resolution is unified to 224×224, normalization is performed, and it is converted to Lab color space to enhance edge features. The audio data was analyzed for time and frequency using Short Time Fourier Transform (STFT) and silence segments were removed. Video data is processed to extract key frames, and motion information is calculated using optical flow. Step 1.3: Text data feature extraction, using TF-IDF and BERT embedding methods to obtain text semantic features; TF-IDF calculation characteristics: ； in, For words In the document The number of times it appears in Total number of documents For inclusion The number of documents; Step 1.4: Image data feature extraction, using wavelet transform (DWT) and ResNet-50 for feature extraction; Wavelet transform characteristics: Perform multi-scale wavelet decomposition on the image to extract low-frequency components. and high frequency components : ； in, For the input image, These are wavelet basis functions; Step 1.5: Audio data feature extraction, calculation of Mel spectral coefficients (MFCC) and spectral centroid; MFCC calculation: ； in, The spectrum is after Mel filtering. For coefficient index; Spectral centroid calculation: ； in, For frequency The power spectrum at that location; Step 1.6: Video data feature extraction. Optical flow method is used to calculate inter-frame motion information and extract spatiotemporal attention features. Features of optical flow calculation: ； in, They represent optical flow at and directional components, Image pixel intensity; Step 1.7: Feature Normalization and Dimensionality Reduction Normalization: The Z-score standardization method is used to make the feature distribution mean 0 and variance 1. ； in, The mean, Standard deviation; Principal component analysis is then used to reduce feature dimensions and improve computational efficiency.

3. The multimodal data security analysis method for the deep fuzzy wavelet learning model according to claim 1, characterized in that: Step 2 involves constructing a deep fuzzy wavelet learning model. Information gain theory is used to filter important features to avoid interference from invalid features in the model's learning. Wavelet transform is incorporated during fuzzy inference to enhance feature representation capabilities, and an adaptive optimization algorithm is employed to optimize the model parameters. The specific steps are as follows: Step 2.1: Feature selection based on information gain The features extracted from the MIDAS and AVID datasets are filtered, the information gain of each feature is calculated, and redundant or invalid features are removed. The formula for calculating information gain is as follows: ； in, Representation of features For target variable Information gain It is the entropy of the target variable. In features Given the circumstances Conditional entropy; Step 2.2: Constructing a fuzzy wavelet neural network The feature space is partitioned using fuzzy C-means (FCM) clustering, and fuzzy membership degrees are calculated: ； in, For the sample Cluster membership degree As the cluster center, For fuzzy index; Then, wavelet transform is used for data feature enhancement, and the obtained fuzzy features are decomposed into multi-scale features: ； in, These are wavelet coefficients. For wavelet basis functions, These represent the scale and translation parameters, respectively. Step 2.3: Optimize the fuzzy rules and model parameters using an adaptive optimization algorithm. The adaptive optimization algorithm is used to optimize fuzzy rules and network parameters to improve the model's learning efficiency and adaptability. This includes the following steps: Adaptive rule adjustment: A data-driven approach is used to dynamically adjust the fuzzy rules in the fuzzy wavelet neural network to adapt to the distribution changes of different categories, with the importance value adjusted from... express: ； in, For the sample On the rules membership degree This is the information gain of the feature; Next, the Adam optimization algorithm is used to optimize the adaptive learning rate: Adam optimization algorithm: ； in, These are the current model parameters. For loss function, The initial learning rate, and First-order and second-order moment estimates, respectively. , The exponential decay rate; Next, the learning rate is dynamically adjusted based on changes in loss to enhance adaptability to multimodal data and perform effective feature selection. The adjustment formula is as follows: ； in, The current learning rate, This is the attenuation coefficient.

4. The multimodal data security analysis method for the deep fuzzy wavelet learning model according to claim 1, characterized in that: Step 3 describes an optimization process that employs a spatial search optimization algorithm. Through global search, local search, and adversarial search strategies, it optimizes model parameters and feature representations to improve generalization ability and adversarial robustness. The specific steps are as follows: Step 3.1: Using the initial model trained in Step 2 as a baseline, construct an optimization population, initialize multiple possible parameter combinations, and ensure that the optimization search range covers a sufficient feature space. Assume the model parameter vector is: ； in, The representative model's first One parameter, initializing the population. As a starting point for optimization, among which Population size; Step 3.2: Use a global search strategy to randomly sample in the high-dimensional feature space, with the goal of minimizing the classification error loss function. ； in, As training samples, For training data, For real labels For the model's predicted output, The cross-entropy loss function is optimized using a gradient estimation-based search strategy. ； in, For learning rate, This represents the gradient of the loss function with respect to the parameters; Step 3.3: Based on the excellent individuals obtained from the global search, a local search is introduced, using gradient information or heuristic methods to fine-tune the model parameters to reduce randomness in the optimization process and improve classification stability; dynamic gradient descent is used to accelerate convergence. ； The momentum factor makes the parameter updates affected by the gradient of the previous time step, thus smoothing the optimization path. Step 3.4: Introduce an adversarial search strategy, construct adversarial features or data augmentation strategies, conduct robustness tests on the model, and optimize the adversarial loss to improve the model's ability to detect abnormal and attack samples. To improve the model's ability to detect adversarial attacks, adversarial perturbations are added. To generate adversarial examples: ； in, For the amplitude of the disturbance, Indicates the gradient direction, generated This can be used to evaluate the adversarial robustness of the model; the optimization objective becomes: ； in, The operation aims to find the most difficult-to-detect adversarial examples. Robustness of the operation optimization model to it; Step 3.5: Evaluate the performance of the optimized model, comparing classification accuracy, robustness metrics, and computational cost before and after optimization. Finally, select the optimal parameter combination to complete the optimization process. Evaluation metrics are as follows: Classification accuracy: ； in, For indicator functions, This is a predicted value; Adversarial example detection success rate: ； in, For adversarial examples The prediction results Reflects the robustness of the model; The final choice is the one that achieves the highest classification accuracy. Minimum optimization parameters As the final model parameters.

5. The multimodal data security analysis method for the deep fuzzy wavelet learning model according to claim 1, characterized in that: Step 4 outputs the security analysis results, which involves detecting the input multimodal data and outputting the data content security analysis results.

Images