Biological heterogeneous data security monitoring method and system based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence and deep learning, and discloses a biological heterogeneous data safety monitoring method and system based on deep learning, and the method comprises the steps: obtaining three types of heterogeneous data: genome, electronic health record and medical image; respectively preprocessing to generate standardized features; the multi-modal alignment encoder is fused into a unified embedded representation; calculating a disturbance sensitivity score by synchronously utilizing a disturbance-resisting sensing unit; triggering a dynamic credibility verification module based on the score, and executing gradient consistency check and feature reconstruction residual analysis; and if detection is abnormal, inference is terminated and an alarm is given, otherwise, a disease prediction backbone network output result is input. The system comprises corresponding functional units, is integrated on a private cloud platform, and supports encrypted transmission and trusted execution environments. According to the invention, cross-modal adversarial attacks can be identified, and safe and reliable operation of medical AI is guaranteed.

Description

A Deep Learning-Based Method and System for Safe Monitoring of Heterogeneous Biological Data Technical Field

[0001] This invention belongs to the field of artificial intelligence and deep learning, specifically relating to a method and system for monitoring the security of heterogeneous biological data based on deep learning. Background Technology

[0002] With the deep penetration of artificial intelligence technology into the healthcare field, deep learning-based biological data analysis systems have been widely applied in key scenarios such as gene sequence analysis, disease risk prediction, and personalized treatment decision support. These systems rely on high-dimensional, heterogeneous biological data to build complex models for accurate modeling and inference of life phenomena. However, the inherent high sensitivity and non-renewability of biological data make it a high-value target for malicious attacks. Especially in open or networked deployment environments, AI medical diagnostic systems face a severe threat from adversarial attacks—attackers can induce models to produce erroneous or even dangerous diagnostic conclusions by injecting minute perturbations imperceptible to the human eye into the input data, seriously jeopardizing patient safety and public trust in healthcare.

[0003] Adversarial attacks targeting heterogeneous biological data exhibit high domain specificity: unlike perturbation patterns in general image or text domains, adversarial noise in biological data must meet strict biological constraints, making it difficult for traditional general-purpose defense mechanisms to effectively identify or defend against such carefully crafted attacks. Existing technologies largely focus on enhancing the robustness of general machine learning models, lacking the ability to model the intrinsic structure, semantic logic, and physiological rationality of biological data, resulting in defense failure when faced with adversarial examples that conform to biological laws. Furthermore, current detection methods often process single-modal data in isolation, failing to fully utilize the inherent correlations between multi-source biological information for cross-validation, making it difficult to accurately identify highly concealed, cross-modal, and collaborative complex attacks in the early stages.

[0004] Existing technologies suffer from multiple shortcomings in addressing these challenges: They lack a dedicated framework capable of simulating the adversarial attack generation process in real-world biological scenarios, resulting in a lack of targeted training for defense strategies; existing detection mechanisms mostly employ static thresholds or fixed rules, failing to adaptively adjust discrimination boundaries based on the dynamic evolution of attack features. More critically, most solutions do not deeply integrate adversarial detection with the model inference process, leading to high detection latency and computational overhead, making it difficult to meet the stringent timeliness requirements of real-time clinical diagnosis. In high-risk applications such as screening for major diseases or gene editing-assisted decision-making, these technological shortcomings can lead to misdiagnosis, missed diagnosis, or even irreversible medical accidents. Therefore, there is an urgent need for an adversarial attack security monitoring and defense system that deeply integrates biological data characteristics, possesses adaptive evolution capabilities, and can be embedded into the diagnostic process. Summary of the Invention

[0005] This invention provides a method and system for monitoring the security of heterogeneous biological data based on deep learning, aiming to address the insufficient robustness and lack of reliability of existing AI medical diagnostic systems when facing adversarial attacks. This method constructs a unified representation space for multi-source heterogeneous biological data, introduces an adversarial perturbation perception mechanism and a dynamic trust verification module, and performs multi-level anomaly detection and real-time verification of model decision paths on input data. This effectively identifies and blocks maliciously constructed adversarial samples during the fusion analysis of multimodal data such as gene sequences, clinical indicators, and imaging features, ensuring the safe operation of deep learning models in real-world medical scenarios.

[0006] This invention provides a deep learning-based method for monitoring the security of biological heterogeneous data, comprising: acquiring biological heterogeneous data to be analyzed, wherein the biological heterogeneous data includes genome sequencing data, electronic health record data, and medical imaging data; performing modality-specific preprocessing on the biological heterogeneous data to generate standardized gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors; inputting the gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors into a multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation; and simultaneously... The adversarial perturbation sensing unit is activated to calculate the statistical deviation of each modal input from the normal distribution and output a perturbation sensitivity score. Based on the perturbation sensitivity score, the dynamic trust verification module is triggered to perform gradient consistency testing and feature reconstruction residual analysis on the multimodal fusion embedded representation. If the gradient consistency test result is less than a preset threshold or the feature reconstruction residual is greater than a set tolerance, it is determined that the current input has an adversarial attack risk, the subsequent diagnostic inference process is terminated, and a security alarm is output. If no security alarm is triggered, the multimodal fusion embedded representation is input into the disease prediction backbone network to generate a disease risk prediction result.

[0007] Preferably, the biological heterogeneous data undergoes modality-specific preprocessing to generate standardized gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors, including: performing quality control filtering on the genome sequencing data to remove reads with a base quality score less than 20; mapping the filtered reads to human reference genome version 38 using a reference genome alignment algorithm; extracting single nucleotide polymorphism sites, insertion / deletion variants, and copy number variation information; encoding the above variant information into a fixed-length one-dimensional binary vector to form the gene feature vector; extracting patient age, gender, past medical history, and laboratory test indicators from the structured fields of the electronic health record data; and performing command processing on the unstructured text. Named entity recognition is used to extract symptom descriptions, medication records, and surgical history; time-stamp-aligned test indicators are resampled at daily granularity to form a numerical matrix; missing values ​​are filled using a forward imputation combined with mean interpolation strategy; all numerical features are standardized with zero mean and unit variance to generate the clinical time-series feature matrix; window width and window level adjustments are performed on the medical image data to adapt to specific tissue types; three-dimensional isotropic resampling is performed to unify voxel spacing; a segmentation model based on a U-shaped convolutional neural network is used to extract the anatomical region of interest; the voxel intensity within the segmentation mask is normalized; the processed three-dimensional image blocks are input into a pre-trained three-dimensional residual network to extract high-level semantic features, forming the image semantic feature tensor.

[0008] Preferably, the gene feature vector, clinical time-series feature matrix, and image semantic feature tensor are input into a multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation, including: mapping the gene feature vector to a 512-dimensional latent space via a fully connected coding sub-network; encoding the clinical time-series feature matrix into a 512-dimensional context vector via a bidirectional long short-term memory network; reducing the dimension of the image semantic feature tensor to 512 dimensions via a projection layer; concatenating the above three 512-dimensional vectors and inputting them into a cross-modal attention fusion layer, calculating the cross-correlation weights among the three, and outputting a 768-dimensional multimodal fusion embedding representation after weighted summation.

[0009] Preferably, during the generation of the multimodal fusion embedding representation, the adversarial perturbation sensing unit is activated simultaneously to calculate the statistical deviation of each modal input relative to the normal distribution and output a perturbation sensitivity score, including: measuring the deviation of the current gene feature vector, clinical time-series feature matrix, and image semantic feature tensor from their respective empirical distribution centers in the training set based on Mahalanobis distance; the empirical distribution centers and covariance matrices are obtained offline through attack-free samples during the model training phase; the perturbation sensitivity score is obtained by weighting the three Mahalanobis distances according to the attack vulnerability index of each modality on the historical validation set.

[0010] Preferably, based on the perturbation sensitivity score, a dynamic reliability verification module is triggered to perform gradient consistency testing and feature reconstruction residual analysis on the multimodal fusion embedding representation, including: calculating the Jacobian matrix of the multimodal fusion embedding representation for each original input modality; performing singular value decomposition on each Jacobian matrix and taking the ratio of the maximum singular value to the minimum singular value as the condition number; if any condition number is greater than 10, the gradient direction is determined to be abnormal; inputting the multimodal fusion embedding representation into the autoencoder structure and outputting the reconstructed original modal features; calculating the mean square error between the reconstructed features and the original input features; if the mean square error of any modality is greater than 0.05, an irreversible perturbation is determined to exist.

[0011] Preferably, if the gradient consistency test result is less than a preset threshold or the feature reconstruction residual is greater than a set tolerance, it is determined that the current input has an adversarial attack risk, the subsequent diagnostic reasoning process is terminated and a security alarm is output, including: determining whether the perturbation sensitivity score is greater than 0.85; determining whether the condition number is greater than 10 or whether the mean square error is greater than 0.05; if any judgment result is yes, a security alarm is triggered and the diagnostic reasoning process is terminated.

[0012] Preferably, if no security alarm is triggered, the multimodal fusion embedding representation is input into the disease prediction backbone network to generate a disease risk prediction result, including: inputting the multimodal fusion embedding representation with superimposed position encoding into a stacked Transformer architecture, the architecture containing six encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network; taking the output vector corresponding to the [CLS] label, mapping it to the probability distribution of a predefined disease category through a fully connected layer to generate the disease risk prediction result.

[0013] Preferably, the autoencoder structure comprises an encoder and a decoder; the encoder consists of three fully connected layers with output dimensions of 768, 512, and 256 respectively; the decoder is symmetrically configured, and its output dimensions are restored to the original dimensions of the gene feature vector, the clinical time-series feature matrix, and the image semantic feature tensor respectively; the loss function adopts a weighted combination of mean squared errors, and the weights are dynamically allocated according to the information entropy of each modality.

[0014] Preferably, the cross-modal attention fusion layer adopts a scaled dot product attention mechanism; the query vector is generated from clinical time-series features, and the key vector and value vector are obtained by projecting gene features and image features, respectively; the attention weights are normalized by the Softmax function and then used to weight image and gene features; the final output is a concatenated vector of clinical features and weighted fusion features.

[0015] This invention also provides a deep learning-based biological heterogeneous data security monitoring system, comprising: a biological heterogeneous data acquisition unit for acquiring biological heterogeneous data to be analyzed, wherein the biological heterogeneous data includes genome sequencing data, electronic health record data, and medical imaging data; a modality-specific preprocessing unit for performing modality-specific preprocessing on the biological heterogeneous data to generate standardized gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors; and a multimodal alignment encoding unit for inputting the gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors into a multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation. The system includes: a perturbation perception unit, used to simultaneously calculate the statistical deviation of each modal input from the normal distribution during the generation of the multimodal fusion embedding representation, and output a perturbation sensitivity score; a dynamic reliability verification unit, used to perform gradient consistency testing and feature reconstruction residual analysis on the multimodal fusion embedding representation based on the perturbation sensitivity score; a safety decision control unit, used to decide whether to terminate the diagnostic inference process and output a safety alarm based on the output of the dynamic reliability verification unit; and a disease prediction backbone network unit, used to input the multimodal fusion embedding representation into the disease prediction backbone network to generate a disease risk prediction result if no safety alarm is triggered.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention achieves comprehensive protection of the input end of the AI ​​medical diagnostic system by constructing a security monitoring system covering three core medical data modalities: genes, clinical and imaging.

[0017] 2. The multimodal alignment coding and adversarial perturbation perception linkage mechanism proposed in this invention can identify distribution anomalies in the early stage of feature fusion, thus avoiding the amplification of malicious perturbations in deep networks.

[0018] 3. The dynamic credibility verification module introduces a dual verification standard of gradient consistency and feature invertibility, which verifies the credibility of the data from two dimensions: the internal behavior of the model and the relationship between input and output, thereby improving the detection accuracy. Attached Figure Description

[0019] Figure 1 is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 is a schematic diagram of the core principle framework of the linkage mechanism between anti-disturbance perception and dynamic trust verification in the present invention; Figure 3 is a flowchart of the modality-specific preprocessing and standardization process for biological heterogeneous data in the present invention; Figure 4 is a flowchart of the logical process of multimodal alignment coding and cross-modal attention fusion in the present invention; Figure 5 is a schematic diagram of the dual-path verification framework of gradient consistency test and feature reconstruction residual analysis of the dynamic trust verification module in the present invention; Figure 6 is a schematic diagram of the multi-level interaction relationship and data flow between the biological heterogeneous data security monitoring system and various platforms of the hospital information system in the present invention. Detailed Implementation

[0020] Please refer to Figures 1 to 6. This invention provides a method and system for monitoring the security of heterogeneous biological data based on deep learning. It aims to prevent malicious attacks against artificial intelligence-based medical diagnostic systems and ensure the robustness and reliability of deep learning models in adversarial environments during the fusion analysis of multimodal data such as gene sequences, clinical indicators, and medical images. This method constructs a unified representation space, introduces an adversarial perturbation perception mechanism and a dynamic trustworthy verification module, and achieves multi-level anomaly detection of input data and real-time verification of model decision paths, thereby effectively identifying and blocking maliciously constructed adversarial samples.

[0021] The method includes the following steps: acquiring heterogeneous biological data to be analyzed; performing modality-specific preprocessing on the heterogeneous biological data; inputting the preprocessed features into a multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation; synchronously activating an adversarial perturbation sensing unit during the fusion process to calculate a perturbation sensitivity score; triggering a dynamic trustworthy verification module based on the score to perform gradient consistency verification and feature reconstruction residual analysis; if any verification index exceeds a set threshold, determining that there is an adversarial attack risk and terminating the subsequent inference process; if no alarm is triggered, inputting the fusion embedding representation into the disease prediction backbone network to generate disease risk prediction results.

[0022] S1. Obtain the biological heterogeneous data to be analyzed, including genome sequencing data, electronic health record data, and medical imaging data.

[0023] Genome sequencing data comes from high-throughput sequencing platforms and stores raw reads in FASTQ format, containing quality score information for each base; electronic health record data is exported from the hospital information system, covering structured fields such as age, gender, and laboratory test values, as well as unstructured text such as outpatient complaints and discharge summaries; medical imaging data is received in digital imaging and communication formats, originating from computed tomography, magnetic resonance imaging, or ultrasound equipment, and includes three-dimensional volume data and corresponding metadata such as scan parameters and patient identifiers.

[0024] All data is accessed through the application programming interface of the hospital's internal private cloud platform to the security monitoring system. The transmission process uses the national cryptographic SM4 algorithm for encryption to ensure the confidentiality and integrity of the data in the transmission link.

[0025] S2. Perform modality-specific preprocessing on the biological heterogeneous data to generate standardized gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors.

[0026] The preprocessing of genome sequencing data first involves quality control filtering, removing reads with a base quality score less than 20 and retaining high-quality reads for subsequent analysis. Then, the BWA-MEM algorithm is used to map the filtered reads to the Human Reference Genome Version 38, generating SAM format alignment files. Based on this, the GATK toolset is used to perform a variant identification process, extracting single nucleotide polymorphism (SNP) sites, insertion / deletion variants, and copy number variations. Finally, this variant information is encoded into a fixed-length one-dimensional binary vector of 100,000 bits, with each bit corresponding to the state of a known pathogenic variant site. If a variant exists at that site, it is set to 1; otherwise, it is set to 0, forming a gene feature vector.

[0027] The preprocessing of electronic health record data first involves parsing structured fields to extract patient age, gender, past medical history, and laboratory test indicators. For unstructured text, named entity recognition is performed, and a BERT-based pre-trained model for the medical field is used to extract symptom descriptions, medication records, and surgical history. Timestamp-aligned test indicators are resampled at a daily granularity to form a numerical matrix with time as rows and indicators as columns, where the number of rows equals the number of observation days and the number of columns equals the number of indicator types. Missing values ​​are filled using a forward imputation combined with mean imputation strategy, prioritizing the use of values ​​from the previous day; if more than three consecutive days of missing values ​​are found, the mean of the corresponding indicator in the training set is used for imputation. Finally, all numerical features are standardized with zero mean and unit variance to generate a clinical time-series feature matrix.

[0028] The preprocessing of medical image data first involves receiving images in digital imaging and communication formats, adjusting the window width and level according to the target tissue type (e.g., a window width of 1500 and a window level of -600 for lung CT); performing three-dimensional isotropic resampling to unify the voxel spacing to 1 mm, and using linear interpolation to ensure geometric consistency; using a pre-trained U-shaped convolutional neural network segmentation model to extract anatomical regions of interest, such as the liver, tumors, or coronary arteries; normalizing the voxel intensity within the segmentation mask to map the original intensity values ​​to the 0-1 range; and inputting the processed three-dimensional image blocks into a pre-trained three-dimensional residual network to extract high-level semantic features, outputting a 768-dimensional vector as the image semantic feature tensor.

[0029] S3. Input the gene feature vector, clinical time-series feature matrix, and image semantic feature tensor into the multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation. The multimodal alignment encoder consists of three independent modality coding sub-networks and one cross-modal attention fusion layer. The gene feature vector is mapped via a fully connected coding sub-network, which contains two fully connected layers. The first layer outputs a dimension of 1024, and the second layer is compressed to 512 dimensions. The activation function used is GeLU. The clinical time-series feature matrix is ​​encoded via a bidirectional long short-term memory network with 256 hidden units. The forward and backward outputs are concatenated and mapped to a 512-dimensional context vector through a linear projection layer. The image semantic feature tensor is reduced to 512 dimensions through a single projection layer. The three are concatenated in 512-dimensional space to form a 1536-dimensional intermediate representation. The cross-modal attention fusion layer uses a scaled dot product attention mechanism. The query vector is generated from the clinical time-series features, and the key vector and value vector are obtained by projecting the gene features and image features, respectively. The attention weight calculation formula is: ;in For clinical feature projection, These are key-value pairs representing gene and image features, respectively. The scaling factor is used; the attention weights are normalized by Softmax and then used to weight the image and gene features; the final output is a concatenated vector of clinical features and weighted fusion features, which is then linearly transformed to 768 dimensions to form a multimodal fusion embedding representation.

[0030] S4. During the generation of the multimodal fusion embedding representation, an adversarial perturbation sensing unit is simultaneously activated to calculate the statistical deviation of each modal input relative to the normal distribution and output a perturbation sensitivity score. The adversarial perturbation sensing unit comprises three parallel distribution offset detectors, respectively targeting genetic, clinical, and time-series image features. Each detector measures the deviation of the current input feature from the empirical distribution center of the training set based on Mahalanobis distance. and covariance matrix The model training phase is achieved through offline estimation using non-attack samples, where... Mahalanobis distance The calculation formula is: ;in For the currently entered number Modal characteristics. The perturbation sensitivity score is a weighted average of three Mahalanobis distances. : Weight The vulnerability index of each modality on the historical validation set is determined. This index is defined as the proportion of misclassifications caused by the modality under standard adversarial attacks, and after normalization, it satisfies the following conditions: Typical values ​​are .

[0031] S5. Based on the perturbation sensitivity score, the dynamic trust verification module is triggered to perform gradient consistency checks and feature reconstruction residual analysis on the multimodal fusion embedding representation. The dynamic trust verification module includes a gradient consistency check unit and a feature reconstruction residual analysis unit. The gradient consistency check unit calculates the multimodal fusion embedding representation. For each original input mode Jacobian matrix Singular value decomposition is performed on each Jacobian matrix to obtain a sequence of singular values. Calculate the condition number .

[0032] If the condition number of any modality is greater than 10, the gradient direction is considered abnormal, indicating that a small input perturbation causes a drastic change in the embedding space, which is consistent with the characteristics of adversarial examples. The feature reconstruction residual analysis unit contains an autoencoder structure, which consists of three fully connected layers with an input dimension of 768 and hidden layer dimensions of 512 and 256 respectively. The decoder is symmetrically set, and the output dimensions are restored to the following dimensions: gene feature vector length of 100,000, clinical time series feature matrix dimension (assuming a maximum observation period of 365 days and 50 indicators, so 18250 dimensions), and image semantic feature tensor dimension of 768. The autoencoder loss function uses a weighted combination of mean squared errors. : ; where weight The information entropy of each modality is dynamically allocated, with higher information entropy resulting in higher weights, ensuring priority is given to reconstructing high-information modalities. After training, the autoencoder parameters are fixed, and the current fused embedding representation is input. Output reconstructed features Calculate the mean square error (MSE) for each mode. If the MSE for any mode is greater than 0.5, it is determined that an irreversible perturbation exists, and the original input may have been maliciously tampered with. For the currently entered number Modal features, For the currently entered number Modal features, For the currently entered number Modal characteristics.

[0033] S6. If the gradient consistency test result is less than a preset threshold or the feature reconstruction residual is greater than a set tolerance, the current input is determined to have a risk of adversarial attack, the subsequent diagnostic reasoning process is terminated, and a security alarm is output. The security decision control unit is set with two levels of judgment logic: the first level judges whether the disturbance sensitivity score is greater than 0.85; the second level judges whether the gradient condition number is greater than 10 or whether the reconstruction residual is greater than 0.05. A security alarm is triggered when either condition is met. The security alarm signal is transmitted to the user interface through the system bus, displaying the prompt "Potential adversarial attack detected, diagnostic process has been terminated", and is simultaneously written to the log audit module, recording the timestamp, patient identifier, statistical quantities of each modality feature, and judgment basis for subsequent review by security analysts.

[0034] S7. If no security alarm is triggered, the multimodal fusion embedding representation is input into the disease prediction backbone network to generate a disease risk prediction result. The disease prediction backbone network is a stacked Transformer architecture containing six encoder layers. Each encoder layer contains a multi-head self-attention mechanism and a feedforward neural network.

[0035] The input embedding is first superimposed with learnable positional encodings, the positional encoding dimension being the same as the embedding dimension, 768; the multi-head attention head is set to 12, each head with a dimension of 64; the feedforward neural network contains two fully connected layers, the intermediate dimension is 3072, and the activation function is GeLU; residual connections and layer normalization are applied to each sub-layer. The final output is the vector corresponding to the [CLS] label, which is mapped to the probability distribution of a predefined disease category through a fully connected layer. The number of categories is set according to the actual task, such as the cancer early screening task including benign, early malignant, and late malignant. The backbone network adopts an adversarial training strategy during the training phase, alternately optimizing the standard cross-entropy loss and the worst-case adversarial perturbation loss. The latter generates the worst perturbation in the embedding space and maximizes the classification loss through the projective gradient descent method, thereby improving the model's generalization ability under perturbation.

[0036] The system includes a biological heterogeneous data acquisition unit, a modality-specific preprocessing unit, a multimodal alignment coding unit, an adversarial disturbance sensing unit, a dynamic trustworthy verification unit, a security decision control unit, and a disease prediction backbone network unit.

[0037] The biological heterogeneous data acquisition unit interfaces with electronic medical record systems, image archiving and communication systems, and high-throughput sequencing data analysis platforms through standardized application programming interfaces (APIs) to achieve automatic data retrieval and format verification. The modality-specific preprocessing unit includes gene preprocessing subunits, clinical preprocessing subunits, and image preprocessing subunits. The modules within each subunit are connected in a pipeline manner to ensure that the processing sequence is irreversible and the status is traceable.

[0038] The multimodal aligned coding unit is deployed on GPU-accelerated computing nodes, supporting batch parallel processing with a throughput greater than 20 instances per second. The adversarial disturbance sensing unit and the dynamic trusted verification unit share the same memory space, avoiding data copying overhead. Their computation graph is jointly compiled with the backbone network, and the Jacobian matrix is ​​efficiently calculated using an automatic differentiation framework.

[0039] The security decision control unit runs on an independent security coprocessor with hardware-level isolation capabilities, ensuring that the decision-making logic is not tampered with by the main system process. The entire system is deployed on the hospital's information system's private cloud platform, running in a trusted execution environment. Memory data is encrypted throughout the process, preventing external processes from accessing model parameters or intermediate features. The system supports a hot update mechanism; when a new attack pattern is identified, new distribution centers and covariance matrices can be pushed through a secure channel, enhancing defense capabilities without system downtime.

[0040] In actual operation, for each patient's data processed by the system, the first step is to complete trimodal data acquisition and preprocessing, which takes approximately 45 seconds; multimodal fusion and adversarial perception are executed simultaneously, taking approximately 8 seconds; dynamic trust verification is performed in parallel along two paths, taking approximately 5 seconds; if the verification passes, the backbone network inference takes approximately 3 seconds. The entire process is completed within 60 seconds, meeting the real-time requirements of clinical practice.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the security of heterogeneous biological data based on deep learning, characterized in that, include: Acquire biological heterogeneous data to be analyzed, including genome sequencing data, electronic health record data, and medical imaging data; The heterogeneous biological data are preprocessed in a modality-specific manner to generate standardized gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors. The gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors are then input into a multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation. During the generation of the multimodal fusion embedded representation, the adversarial perturbation sensing unit is activated simultaneously to calculate the statistical deviation of each modal input from the normal distribution and output the perturbation sensitivity score. Based on the perturbation sensitivity score, the dynamic trust verification module is triggered to perform gradient consistency check and feature reconstruction residual analysis on the multimodal fusion embedding representation; if the gradient consistency check result is less than the preset threshold or the feature reconstruction residual is greater than the set tolerance, it is determined that the current input has an adversarial attack risk, the subsequent diagnostic reasoning process is terminated and a security alarm is output. If no security alarm is triggered, the multimodal fusion embedding representation is input into the disease prediction backbone network to generate disease risk prediction results.

2. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 1, characterized in that, Modality-specific preprocessing is performed on the aforementioned heterogeneous biological data to generate standardized gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors. This includes: performing quality control filtering on the genome sequencing data to remove reads with a base quality score less than 20; mapping the filtered reads to Human Reference Genome Version 38 using a reference genome alignment algorithm; extracting single nucleotide polymorphism sites, insertion / deletion variants, and copy number variations; encoding the above variation information into a fixed-length one-dimensional binary vector to form the gene feature vector; extracting patient age, gender, past medical history, and laboratory test indicators from the structured fields of the electronic health record data; and performing naming operations on the unstructured text. The process involves: extracting symptom descriptions, medication records, and surgical history using body recognition; resampling time-stamp-aligned test indicators at a daily granularity to form a numerical matrix; filling missing values ​​using a forward imputation combined with mean interpolation; standardizing all numerical features to zero mean and unit variance to generate the clinical time-series feature matrix; adjusting the window width and level of the medical image data to suit specific tissue types; performing three-dimensional isotropic resampling to unify voxel spacing; extracting anatomical regions of interest using a segmentation model based on a U-shaped convolutional neural network; normalizing the voxel intensity within the segmentation mask; and inputting the processed three-dimensional image blocks into a pre-trained three-dimensional residual network to extract high-level semantic features, forming the image semantic feature tensor.

3. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 2, characterized in that, The gene feature vector, clinical temporal feature matrix, and image semantic feature tensor are input into a multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation. This includes: mapping the gene feature vector to a 512-dimensional latent space via a fully connected encoding subnetwork; encoding the clinical temporal feature matrix into a 512-dimensional context vector via a bidirectional long short-term memory network; reducing the dimension of the image semantic feature tensor to 512 dimensions via a projection layer; concatenating the above three 512-dimensional vectors and inputting them into a cross-modal attention fusion layer; calculating the cross-correlation weights among the three; and outputting a 768-dimensional multimodal fusion embedding representation after weighted summation.

4. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 3, characterized in that, During the generation of the multimodal fusion embedding representation, the adversarial perturbation sensing unit is activated simultaneously to calculate the statistical deviation of each modal input from the normal distribution and output a perturbation sensitivity score. This includes: measuring the deviation of the current gene feature vector, clinical time-series feature matrix, and image semantic feature tensor from their respective empirical distribution centers in the training set based on Mahalanobis distance; the empirical distribution centers and covariance matrices are obtained offline through attack-free samples during the model training phase; and the perturbation sensitivity score is obtained by weighting the three Mahalanobis distances according to the attack vulnerability index of each modality on the historical validation set.

5. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 4, characterized in that, Based on the perturbation sensitivity score, the dynamic reliability verification module is triggered to perform gradient consistency testing and feature reconstruction residual analysis on the multimodal fusion embedding representation, including: calculating the Jacobian matrix of the multimodal fusion embedding representation for each original input modality; performing singular value decomposition on each Jacobian matrix and taking the ratio of the maximum singular value to the minimum singular value as the condition number; if any condition number is greater than 10, the gradient direction is determined to be abnormal; inputting the multimodal fusion embedding representation into the autoencoder structure and outputting the reconstructed original modal features; calculating the mean square error between the reconstructed features and the original input features; if the mean square error of any modality is greater than 0.05, an irreversible perturbation is determined to exist.

6. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 5, characterized in that, If the gradient consistency test result is less than a preset threshold or the feature reconstruction residual is greater than a set tolerance, it is determined that the current input has an adversarial attack risk, the subsequent diagnostic reasoning process is terminated and a security alarm is output, including: determining whether the perturbation sensitivity score is greater than 0.85; determining whether the condition number is greater than 10 or whether the mean square error is greater than 0.05; if any judgment result is yes, a security alarm is triggered and the diagnostic reasoning process is terminated.

7. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 6, characterized in that, If no security alarm is triggered, the multimodal fusion embedding representation is input into the disease prediction backbone network to generate a disease risk prediction result. This includes: inputting the multimodal fusion embedding representation with superimposed position encoding into a stacked Transformer architecture, which contains six encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network; taking the output vector corresponding to the [CLS] label and mapping it to the probability distribution of a predefined disease category through a fully connected layer to generate the disease risk prediction result.

8. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 7, characterized in that, The autoencoder structure comprises an encoder and a decoder. The encoder consists of three fully connected layers with output dimensions of 768, 512, and 256, respectively. The decoder is symmetrically configured, and its output dimensions are restored to the original dimensions of the gene feature vector, the clinical time-series feature matrix, and the image semantic feature tensor, respectively. The loss function is a weighted combination of mean squared errors, with the weights dynamically allocated based on the information entropy of each modality.

9. The method for monitoring the security of heterogeneous biological data based on deep learning according to claim 8, characterized in that, The cross-modal attention fusion layer employs a scaled dot product attention mechanism; The query vector is generated from clinical time-series features, while the key vector and value vector are obtained by projecting gene features and image features, respectively. Attention weights are normalized using the Softmax function and then used to weight image and genetic features; the final output is a concatenated vector of clinical features and weighted fusion features.

10. A deep learning-based biological heterogeneous data security monitoring system, characterized in that, include: A biological heterogeneous data acquisition unit is used to acquire biological heterogeneous data to be analyzed, including genome sequencing data, electronic health record data, and medical imaging data. A modality-specific preprocessing unit is used to perform modality-specific preprocessing on the biological heterogeneous data to generate standardized gene feature vectors, clinical time-series feature matrices, and image semantic feature tensors. A multimodal alignment encoding unit is used to input the gene feature vector, clinical time-series feature matrix, and image semantic feature tensor into a multimodal alignment encoder to generate a unified-dimensional multimodal fusion embedding representation; an adversarial perturbation perception unit is used to simultaneously calculate the statistical deviation of each modality input relative to the normal distribution during the generation of the multimodal fusion embedding representation and output a perturbation sensitivity score; a dynamic reliability verification unit is used to perform gradient consistency testing and feature reconstruction residual analysis on the multimodal fusion embedding representation based on the perturbation sensitivity score. The safety decision control unit is used to determine whether to terminate the diagnostic reasoning process and output a safety alarm based on the output of the dynamic trusted verification unit. The disease prediction backbone network unit is used to input the multimodal fusion embedded representation into the disease prediction backbone network to generate disease risk prediction results without triggering a security alarm.

Images