Data security protection method and data security protection system

Abstract

The invention relates to a data security protection method and a data security protection system, and relates to the technical field of data processing. The data security protection method is applied to the medical information system, and comprises the following steps: acquiring heterogeneous data in the medical information system; the heterogeneous data comprises text data, image data and physiological signal data; inputting the heterogeneous data into a multi-modal fusion model, and outputting a threat representation vector in a preset format; the threat representation vector is used for representing the abnormal degree of the heterogeneous data; performing semantic analysis on the operation instruction by using a large language model, and jointly calculating a risk score according to the threat representation vector; the risk score is used for representing the risk degree of the operation instruction for initiating access to data in the medical information system; and under the condition that the risk score exceeds a preset first threshold value, upgrading the encryption level of the accessed data and triggering an enhanced authentication process. According to the invention, the data security protection performance of the medical information system can be improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a data security protection method and a data security protection system. Background Technology

[0002] With the rapid development of smart healthcare systems, medical data is growing exponentially, encompassing diverse and heterogeneous sources such as electronic medical records, medical images, and physiological signal data from wearable devices. This data plays a crucial role in disease diagnosis, remote monitoring, and AI-assisted treatment; however, its high value also makes it a primary target for cyberattacks. Nevertheless, existing technologies have significant limitations in medical data security, and the risk of data leakage and attack remains extremely high. Summary of the Invention

[0003] Therefore, it is necessary to provide a data security protection method and system to address the aforementioned technical issues, thereby improving the data security protection performance of medical information systems.

[0004] This application provides a data security protection method applied to a medical information system, the method comprising:

[0005] Acquire heterogeneous data from the medical information system; the heterogeneous data includes text data, image data, and physiological signal data;

[0006] The heterogeneous data is input into a multimodal fusion model, which outputs a threat characterization vector in a preset format; the threat characterization vector is used to represent the degree of anomaly of the heterogeneous data.

[0007] A large language model is used to perform semantic parsing on the operation instructions, and a risk score is jointly calculated based on the threat representation vector; the risk score is used to represent the degree of risk of the operation instructions initiating access to data in the medical information system.

[0008] If the risk score exceeds a preset first threshold, the encryption level of the accessed data is upgraded and an enhanced authentication process is triggered.

[0009] In one embodiment, the multimodal fusion model includes a natural language processing network, a convolutional neural network, a temporal neural network, a cross-modal attention mechanism, and an interactive feature fusion algorithm; the step of inputting the heterogeneous data into the multimodal fusion model and outputting a threat representation vector in a preset format includes:

[0010] The text data is input into the natural language processing network for feature extraction to generate text features;

[0011] The image data is input into the convolutional neural network for feature extraction to generate image features;

[0012] The physiological signal data is input into the temporal neural network for feature extraction to generate physiological features;

[0013] The cross-modal attention mechanism is used to perform spatiotemporal alignment of the text features, image features, and physiological features to generate fused features;

[0014] The interactive feature fusion algorithm is used to mine causal anomalies between different data sources in the fused features and generate a threat characterization vector in a preset format.

[0015] In one embodiment, the step of semantically parsing the operation instructions using a large language model and jointly calculating a risk score based on the threat representation vector includes:

[0016] Acquire the operation instruction and the operation-related data corresponding to the operation instruction;

[0017] The operation instructions and operation-related data are input into the large language model for semantic parsing to generate an operation intent vector.

[0018] The risk score is obtained by jointly calculating the operational intent vector and the threat representation vector.

[0019] In one embodiment, the method further includes:

[0020] If the risk score exceeds a preset first threshold, the operation instructions and the threat representation vector are input into the large language model, and a causal correlation audit report is output.

[0021] The multimodal fusion network is updated based on the audit results corresponding to the causal correlation audit report.

[0022] In one embodiment, the method further includes:

[0023] Obtain threat feature information generated locally by the medical institution based on the threat representation vector;

[0024] The threat signature information is processed for privacy protection to generate privacy-protected data;

[0025] By using a cloud-based federated learning architecture, the privacy-protected data sent by various medical institutions is aggregated and trained to generate a global threat detection model.

[0026] The global threat detection model is lightweighted to obtain a lightweight model, which is then deployed to the local systems and edge devices of various medical institutions.

[0027] In one embodiment, the threat feature information includes a threat feature vector summary and / or parameter update gradients; wherein,

[0028] The step of performing privacy protection processing on the threat feature information to generate privacy-protected data includes:

[0029] Differential privacy technology is used to add noise to the threat feature vector summary and / or parameter update gradient of the medical institution's local data to generate the privacy-protected data.

[0030] In one embodiment, the method further includes:

[0031] Anomaly detection is performed on the image data and text data in the heterogeneous data respectively to generate corresponding image anomaly indicators and text anomaly indicators.

[0032] Based on the combined judgment results of the image anomaly indicators and text anomaly indicators, a judgment result on whether there is an adversarial attack is determined.

[0033] If the judgment result is yes, a security isolation mechanism is triggered and the attack evidence is preserved in an immutable manner;

[0034] The characteristic information of the adversarial attack is sent to the cloud-based federated learning architecture to update the global threat detection model.

[0035] This application provides a data security protection system for use in a medical information system, including:

[0036] A multimodal perception module is used to acquire heterogeneous data in the medical information system; the heterogeneous data includes text data, image data, and physiological signal data.

[0037] The threat representation learning module is used to input the heterogeneous data into a multimodal fusion model and output a threat representation vector in a preset format; the threat representation vector is used to represent the degree of anomaly of the heterogeneous data.

[0038] The dynamic protection module is used to perform semantic parsing of operation instructions using a large language model and jointly calculate a risk score based on the threat representation vector. The risk score is used to represent the risk level of the operation instruction initiating access to data in the medical information system. If the risk score exceeds a preset first threshold, the encryption level of the accessed data is upgraded and an enhanced authentication process is triggered.

[0039] In one embodiment, the system further includes: a federated learning module, used to aggregate privacy-preserving data sent by various medical institutions, train and generate a global threat detection model, perform lightweight processing on the global threat detection model to obtain a lightweight model, and deploy the lightweight model to the local systems and edge devices of various medical institutions; wherein, the privacy-preserving data is threat feature information generated locally by the medical institutions based on the threat representation vector, and the threat feature information is generated by performing privacy-preserving processing.

[0040] In one embodiment, the system further includes: an adversarial attack defense module, used to perform anomaly detection on the image data and text data in the heterogeneous data respectively, and generate corresponding image anomaly indicators and text anomaly indicators; determine whether an adversarial attack exists based on the combined judgment result of the image anomaly indicators and text anomaly indicators; if the judgment result is yes, trigger a security isolation mechanism and perform tamper-proof evidence storage of the attack; and send the feature information of the adversarial attack to the cloud federated learning architecture to update the global threat detection model.

[0041] The aforementioned data security protection methods and systems achieve comprehensive threat perception coverage by acquiring heterogeneous data from medical information systems, including text data, image data, and physiological signal data. This heterogeneous data is then input into a pre-constructed multimodal fusion model. This model extracts features from text data, image data, and physiological signals, performs cross-modal spatiotemporal alignment, and fuses these features to output a threat representation vector in a preset format. This vector accurately represents the degree and type of anomaly in the heterogeneous data. Furthermore, a medical-specific large language model is used to semantically parse medical data access operation instructions, extracting key information from these instructions. This information is then combined with the threat representation vector to calculate a risk score. This risk score quantitatively represents the risk level of the operation instruction's access to data in the medical information system, enabling dynamic adaptation of risk assessment. If the risk score exceeds a preset first threshold, the encryption level of the accessed data is automatically upgraded, triggering an enhanced authentication process. This achieves dynamic adaptive adjustment of medical data security protection, effectively intercepting high-risk access behaviors and ensuring medical data privacy, while avoiding the impact of excessive protection on normal medical business efficiency. It also provides core data support for subsequent iterative optimization of protection strategies and cross-institutional collaborative protection. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating a data security protection method in one embodiment;

[0043] Figure 2 This is a flowchart illustrating a data security protection method in another embodiment;

[0044] Figure 3 This is a flowchart illustrating a data security protection method in yet another embodiment;

[0045] Figure 4 This is a flowchart illustrating a data security protection method in another embodiment;

[0046] Figure 5 This is a structural block diagram of a data security protection system in one embodiment;

[0047] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] Medical data security faces three major challenges: First, the risk of data breaches is increasing, with hackers using social engineering attacks to illegally gain access to doctors' workstations, leading to the leakage of patients' private information. Second, AI adversarial attacks are frequent, with attackers using generative adversarial networks (GANs) or other technologies to tamper with medical image diagnostic results, misleading clinical decisions. Third, collaborative protection of multimodal data is difficult, as text medical records, medical images, and electrocardiogram time-series signals are difficult to protect against with unified security strategies due to their different data formats. Traditional rule-based authentication, static encryption algorithms, and single-modal anomaly detection technologies are no longer sufficient to meet the needs of dynamic threat defense, necessitating the construction of a new security architecture that integrates context awareness, multimodal analysis, and adaptive response.

[0050] In related technologies, medical data security control strategies generally rely on static rules and fixed strategies, making it difficult to cope with new dynamic threats. There is a contradiction between medical data encryption mechanisms and the resource limitations of edge devices. Existing solutions, when implementing high-strength encryption on wearable devices or portable monitors, often cause data transmission delays due to excessive computational load, affecting clinical treatment efficiency. Furthermore, the ability to perceive medical threats is also limited by the isolated nature of data processing. Existing systems lack a unified analysis framework for multi-source heterogeneous data such as text medical records, medical images, and electrocardiogram signals, leading to spatiotemporal gaps in attack chain detection. Traditional privacy protection technologies may also disrupt data correlation during data anonymization, affecting both the training effect of AI-assisted diagnostic models and failing to meet the requirements of a unified security strategy for multimodal data, creating a dual dilemma of privacy protection and clinical value extraction.

[0051] In response, this application provides a data security protection method, device, computer equipment, medium, and program product, which overcomes the limitations of single-modal detection in related technologies, enhances threat identification capabilities, and improves the data security protection performance of the system.

[0052] In some exemplary embodiments, such as Figure 1 As shown, a data security protection method is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the data security protection method is applied to a medical information system, and the method includes the following steps S102 to S108.

[0053] S102, acquire heterogeneous data from the medical information system; heterogeneous data includes text data, image data and physiological signal data.

[0054] Medical information systems refer to integrated information platforms that support clinical diagnosis and treatment, data management, and cross-institutional collaboration. They include core subsystems such as Electronic Medical Record System (EMR), Picture Archiving and Communication Systems (PACS), Laboratory Information System (LIS), and physiological signal monitoring system, covering various application scenarios such as hospitals, health check-up centers, and regional medical collaboration platforms.

[0055] Text data, using natural language as its medium, is structured or unstructured data that records semantic information such as medical diagnosis and treatment processes, patient information, and medical decisions. It is the core carrier of logical connections and diagnostic and treatment intentions in medical data, including but not limited to electronic medical records, medical orders, diagnostic reports, and examination request forms. Imaging data, acquired through medical imaging equipment (such as CT and MRI scanners), is visualized data reflecting human anatomical structures, pathological features, or functional states, using pixels / voxels as the basic unit. It is the core carrier of intuitive morphological and lesion information in medical data, including but not limited to CT (Computed Tomography), MRI (Magnetic Resonance Imaging), ultrasound, and pathological slides. Physiological signal data, acquired in real time through sensors, is a temporal electrical and physical signal reflecting dynamic changes in human physiological functions. It is the core carrier of dynamic physiological states and abnormal trends in medical data, including but not limited to temporal physiological data such as electrocardiograms (ECG), electroencephalograms (EEG), blood oxygen saturation, and blood pressure.

[0056] In applications, a multimodal perception module can be deployed at the data acquisition end to simultaneously collect heterogeneous data from multiple sources, including text data, medical imaging data, and physiological signal data. This multimodal perception module can be deployed at various clinical data-generating nodes within the hospital, including core scenarios such as examination rooms, examination rooms, intensive care units, and medical laboratories. The module can adopt a modular design, integrating data acquisition, format conversion, clock synchronization, and secure transmission modules, and possesses multi-interface adaptability to be compatible with different types of medical equipment and information systems. Furthermore, a multi-channel parallel acquisition mechanism can be employed, with dedicated acquisition processes designed for text data, medical imaging data, and physiological signal data, while simultaneously achieving synchronous association of the three types of data through timestamp binding. The collected data can be transmitted to the threat characterization module via a dedicated isolated hospital network.

[0057] S104: Input heterogeneous data into the multimodal fusion model and output a threat characterization vector in a preset format; the threat characterization vector is used to represent the degree of anomaly of the heterogeneous data.

[0058] The multimodal fusion model is pre-built and trained, and can be dynamically updated. It refers to a dedicated deep learning network model that adapts to the characteristics of multi-source heterogeneous data in medical information systems, integrating multimodal feature extraction, cross-modal spatiotemporal alignment, and feature fusion functions. Its core design goal is to address the heterogeneity of text, images, and physiological signals in terms of morphology, dimension, and feature type, achieving accurate discovery of collaborative anomalous features in multi-source heterogeneous data. This multimodal fusion model is not a single network structure, but an integrated model composed of multiple networks. In application, the multimodal fusion model can perform feature extraction, cross-modal spatiotemporal alignment, and feature fusion processing on multi-source heterogeneous data, and output a threat representation vector.

[0059] A threat characterization vector (HMV) is a fixed-dimensional vector data output by a multimodal fusion model, used to quantify the overall anomaly level of multi-source heterogeneous medical data. Essentially, it transforms anomalous information in multimodal data (such as text tampering traces, image forgery features, and abnormal fluctuations in physiological signals) into a machine-recognizable and calculable numerical form. Specifically, the preset format can be a fixed-dimensional real-number vector, with the number of dimensions preset according to the anomaly type coverage requirements for medical data security protection. Each element in the vector corresponds to the quantification level of a type of anomaly; the larger the element value, the higher the confidence level of the corresponding type of anomaly. The core function of this vector is to achieve a unified representation of anomaly information, breaking through the limitations of scattered anomaly features and the inability to conduct collaborative analysis in traditional single-modal detection. Simultaneously, through quantitative representation, it provides clear input for subsequent risk assessment, enabling risk scores to accurately reflect the comprehensive security status of multi-source data.

[0060] In applications, the threat characterization module can extract features and perform cross-modal alignment on the multi-source heterogeneous data collected by the multimodal perception module, and push the generated threat characterization vector to the dynamic protection engine to provide real-time and accurate abnormal feature basis for subsequent risk scoring calibration and protection strategy adjustment.

[0061] S106 utilizes a large language model to perform semantic parsing of operation instructions and jointly calculates a risk score based on threat representation vectors; the risk score is used to represent the degree of risk of the operation instructions initiating access to data in the medical information system.

[0062] In this embodiment, the large language model refers to a specialized language model pre-trained and fine-tuned using medical-specific corpora. Unlike general-purpose large language models, it can accurately parse the semantics of operational instructions in medical scenarios and uncover the deep access intentions behind these instructions, effectively identifying risks that traditional static access control struggles to detect, such as malicious operations by legitimate roles. The risk score is a quantitative value calculated using a weighted fusion algorithm, combining the operational intention vector and threat representation vector obtained from the large language model. This score accurately characterizes the security risk level of operational instructions accessing medical information system data. A higher score indicates a higher probability of unauthorized access, data theft, malicious tampering, or other security threats, providing a clear quantitative basis for triggering subsequent protection strategies.

[0063] In applications, large language models can be deployed on a dynamic protection engine. The engine calculates risk scores and, combined with the output threat representation vectors, dynamically matches and triggers corresponding security protection policies (such as initiating enhanced authentication when the risk score exceeds a threshold). The dynamic protection engine can be deployed on core service nodes of medical information systems to achieve centralized control and policy coordination of access behavior across the entire system. It also adapts to the distributed architecture of hospital intranets, deploying edge nodes in clinical subsystems such as electronic medical record systems, medical image archiving and communication systems, and laboratory information systems to reduce response latency of protection policies and ensure real-time protection performance in clinical scenarios.

[0064] S108: If the risk score exceeds the preset first threshold, upgrade the encryption level of the accessed data and trigger an enhanced authentication process.

[0065] The preset first threshold is a pre-defined risk score threshold used to define the security risk level of access behavior. Its value is determined based on medical data security standards, clinical business scenario needs, and historical risk event data calibration, and can be set between 60 and 80 (corresponding to a risk score range of 0-100). For example, for highly sensitive medical data such as gene testing data and mental illness medical records, the first threshold can be lowered to 60 for stricter risk control; for ordinary medical data such as routine physical examination reports, the first threshold can be raised to 75 to balance security and diagnostic efficiency. This threshold supports dynamic iterative optimization and can be adjusted based on the medical institution's experience in handling risk events and updates to industry security standards to ensure the accuracy of risk assessment.

[0066] The accessed data refers to the medical data directly accessed by the user's operation commands, encompassing various multi-source heterogeneous data collected in step S102 above. It is important to note that the accessed data includes not only the data itself but also related derived information (such as image diagnostic annotations and structured medical record summaries). The protection scope must cover the entire data transmission, storage, and retrieval chain to prevent data leakage or tampering due to omissions in the protection scope. Encryption level refers to the security strength level of encryption protection for medical data. This solution pre-sets multiple encryption strategies, with different levels corresponding to different encryption algorithms, key lengths, and encryption ranges. These can be divided into two categories: basic encryption level and enhanced encryption level. The basic encryption level is suitable for low-risk access scenarios, employing a lightweight encryption algorithm (such as AES-128) to encrypt only the data transmission link, ensuring basic security while reducing system overhead. The enhanced encryption level is suitable for high-risk scenarios where the risk score exceeds a preset first threshold, employing a high-strength encryption algorithm (such as the national cryptographic algorithm SM4), with encryption covering the data transmission link, storage medium, and retrieval process. It also increases the key length (e.g., using a 256-bit key) to significantly improve the data's resistance to cracking.

[0067] Enhanced authentication processes refer to multi-level identity verification procedures initiated for high-risk access behaviors. Unlike conventional single-password authentication, its core objective is to ensure the legitimacy of the access behavior and the authenticity of the operator. Specifically, biometric authentication, such as fingerprint or iris recognition, can be triggered first, comparing pre-entered biometric information from medical personnel. If biometric authentication passes, multi-factor authentication, such as password and dynamic password, can be further initiated. The dynamic password is issued via a dedicated hospital authentication terminal or encrypted SMS. After completing the above authentication, the system can also display a pop-up window explaining the purpose of the access, requiring the operator to fill in the reason for the access and keeping a record, providing traceable evidence for subsequent audits. For cases where authentication fails, the system can directly block the access request and generate a risk alarm log, pushing it to the security management platform.

[0068] In the application, the dynamic protection engine compares the risk score with a preset first threshold in real time. If the risk score exceeds the threshold, an encryption level upgrade command is immediately triggered. The encryption policy of the accessed data is switched from the basic level to the enhanced level via the system interface, and an enhanced authentication request is pushed to the operating terminal. Access to the data is only allowed after the operator completes the enhanced authentication and passes verification. During the access process, the system continuously monitors the operation behavior, such as for abnormal operations like batch downloads or data tampering. If authentication fails or new high-risk behavior is detected during access, the system will immediately block the access link, freeze the access permissions of the corresponding operating account, and generate a risk event report to be submitted to the manual audit module. If the risk score is less than the preset first threshold, the operation behavior is considered normal, and monitoring can continue.

[0069] The data security protection method provided in the above embodiments achieves comprehensive threat perception coverage by acquiring heterogeneous data from the medical information system, including text data, image data, and physiological signal data. The heterogeneous data is then input into a pre-constructed multimodal fusion model. This model extracts features from the text data, image data, and physiological signals, performs cross-modal spatiotemporal alignment, and fuses features to output a threat representation vector in a preset format. This threat representation vector accurately represents the degree and type of anomaly in the heterogeneous data. Furthermore, a medical-specific large language model is used to semantically parse medical data access operation instructions, extracting key information from the instructions. This information is then combined with the threat representation vector to calculate a risk score. The risk score quantitatively represents the risk level of the operation instruction's access to data in the medical information system, achieving dynamic adaptation of risk assessment. If the risk score exceeds a preset first threshold, the encryption level of the accessed data is automatically upgraded, and an enhanced authentication process is triggered. This achieves dynamic adaptive adjustment of medical data security protection, effectively intercepting high-risk access behaviors and ensuring medical data privacy, while avoiding the impact of excessive protection on normal medical business efficiency. It also provides core data support for subsequent iterative optimization of protection strategies and cross-institutional collaborative protection.

[0070] In some exemplary embodiments, such as Figure 2 As shown, the multimodal fusion model includes a natural language processing network, a convolutional neural network, a temporal neural network, a cross-modal attention mechanism, and an interactive feature fusion algorithm. S104 involves inputting heterogeneous data into the multimodal fusion model and outputting a threat representation vector in a preset format, including the following S202 to S210.

[0071] S202, input the text data into the natural language processing network for feature extraction and generate text features.

[0072] Natural Language Processing (NLP) networks are specialized network models adapted for semantic parsing of medical text data. For example, an NLP network can employ a network architecture optimized based on a BioBERT pre-trained model. NLP networks can be fine-tuned by incorporating medical domain corpora, possessing core capabilities for accurate identification of medical terminology and semantic association mining. The network structure may include a Transformer encoder, a medical terminology embedding layer, and a pooling layer, effectively adapting to the semantic features of medical texts such as electronic medical records, medical orders, and examination reports.

[0073] Text features refer to fixed-dimensional semantic vectors output after processing by the aforementioned natural language processing network. Each element in the vector corresponds to the semantic representation of key information in the text data, specifically covering the semantic attributes of core content such as disease diagnosis conclusions, treatment plans, examination indicators, and patient basic information.

[0074] In the application, the acquired medical text data can first be preprocessed, including removing invalid characters, standardizing medical terminology, and defining sentence boundaries. Then, the preprocessed text data can be segmented into segments according to a preset number of units and input into a pre-trained and optimized natural language processing network. The network captures the contextual dependencies of words in the text through a Transformer encoder, transforms the text into vector representations through a medical term embedding layer, and then performs dimensionality reduction and feature aggregation on the vectors through a pooling layer, finally outputting text features. These text features will serve as one of the core inputs for the subsequent step S208 cross-modal feature spatiotemporal alignment, laying the data foundation of the text dimension for the fusion analysis of multi-source heterogeneous data.

[0075] S204: Input the image data into a convolutional neural network for feature extraction and generate image features.

[0076] Convolutional Neural Networks (CNNs) are specialized network models adapted for extracting key regions from medical image data. For example, a CNN can employ an improved network architecture based on ResNet-50. CNNs can be optimized through transfer learning by incorporating labeled medical image datasets, possessing core capabilities for precise localization of lesions and extraction of anatomical features from medical images. The network structure can include residual blocks, convolutional layers, max-pooling layers, and global average pooling layers, effectively adapting to the grayscale and spatial structural features of various types of medical images, such as CT, MRI, and ultrasound. Image features refer to the fixed-dimensional feature vector output after processing by the aforementioned CNN. Each element in the vector corresponds to the spatial and grayscale attributes of key regions in the medical image, including but not limited to the location, size, edge contour, and density distribution of lesions, as well as the morphological features of normal anatomical structures.

[0077] In the application, the acquired medical image data can first be preprocessed, including Gaussian filtering for noise reduction, grayscale normalization, size standardization, and channel expansion. Then, the preprocessed image data can be input into a convolutional neural network optimized by transfer learning. The network extracts low-, medium-, and high-order features of the image layer by layer through the initial convolutional layer and residual blocks, compresses the feature map dimension using the max pooling layer while retaining key information, and then transforms the multi-dimensional feature map into a one-dimensional vector through the global average pooling layer, finally outputting the image features. These image features will serve as one of the core inputs for the subsequent step S208 cross-modal feature spatiotemporal alignment, laying the data foundation of image dimension for the fusion analysis of multi-source heterogeneous data.

[0078] S206, physiological signal data is input into a time-series neural network for feature extraction to generate physiological features.

[0079] Temporal neural networks are specialized network models adapted for capturing the temporal features of medical physiological signals. For example, a temporal neural network employs an improved architecture based on BiLSTM (Bidirectional Long Short-Term Memory). Temporal neural networks can be optimized through transfer learning by introducing labeled datasets of medical physiological signals, possessing the core capability of accurately capturing the dynamic changes in physiological signals such as heart rate and blood oxygen saturation. The network structure may include bidirectional LSTM layers, fully connected layers, and dropout layers, effectively adapting to the feature extraction needs of temporal physiological signals such as electrocardiograms (ECG), blood oxygen saturation (SpO2), and blood pressure waveforms. Physiological features refer to the fixed-dimensional feature vector output after processing by the aforementioned temporal neural network. Each element in the vector corresponds to the temporal dynamic attribute representation of the physiological signal, specifically covering core content such as heart rate change trends, blood oxygen fluctuation amplitude, blood pressure peak intervals, and respiratory rate patterns.

[0080] In the application, the acquired physiological signal data can be preprocessed first, including median filtering for noise reduction, signal normalization, length standardization, and outlier removal. Then, the preprocessed physiological signal sequence is input into the temporal neural network optimized by transfer learning. The network captures the forward temporal dependence and reverse temporal correlation of the signal simultaneously through a bidirectional LSTM layer, aggregates and adjusts the temporal features through a fully connected layer, and then suppresses overfitting through a dropout layer, finally outputting physiological features. These physiological features will serve as one of the core inputs for the subsequent step S208 cross-modal feature spatiotemporal alignment, laying the data foundation for the physiological signal dimension of the fusion analysis of multi-source heterogeneous data.

[0081] S208 utilizes a cross-modal attention mechanism to spatiotemporally align text features, image features, and physiological features to generate fused features.

[0082] Cross-modal attention mechanisms are specialized collaborative mechanisms adapted to align heterogeneous features from multi-source medical data. For example, a cross-modal attention mechanism can be implemented using an improved Transformer-based cross-attention architecture. This mechanism can be trained and optimized by introducing a medical multimodal association annotation dataset, enabling the mining and accurate alignment of spatiotemporal associations among three types of heterogeneous features: text, images, and physiological signals. Its core structure includes a feature mapping layer, a multi-head cross-attention layer, and a feature fusion layer, effectively addressing the spatiotemporal dimensional mismatch caused by differences in data types among the three types of features. The fused feature refers to the fixed-dimensional feature vector output after alignment and aggregation by the aforementioned cross-modal attention mechanism. Each element in the vector represents the spatiotemporal association attributes of the three types of data: text, images, and physiological signals. Specifically, it can cover the spatiotemporal causal relationship between disease diagnosis conclusions and corresponding image lesions and physiological signal abnormalities.

[0083] In the application, the aforementioned text features, image features, and physiological features can be preprocessed to ensure that the three types of features are in the same dimensional space. Then, the preprocessed three types of features can be input into a multi-head cross-attention layer to mine the spatiotemporal correlation of the three types of features by calculating attention weights. Next, the feature fusion layer can perform element-level addition and normalization operations on the weighted three types of features to complete feature aggregation. Finally, the fused feature is output. This fused feature eliminates the spatiotemporal differences and type barriers of multi-source heterogeneous medical data, providing a unified feature input basis for the causal anomaly mining in the subsequent step S210.

[0084] S210 utilizes an interactive feature fusion algorithm to mine causal anomalies between different data sources in the fused features, generating a threat characterization vector in a preset format.

[0085] The interactive feature fusion algorithm is a specialized algorithm adapted for causal anomaly mining of multi-source medical fusion features. For example, the interactive feature fusion algorithm can be implemented using an improved graph neural network architecture based on GraphSAGE (Graph Sample and Aggregated). The interactive feature fusion algorithm can be trained and optimized by introducing a medical multimodal anomaly association annotation dataset, possessing the core capabilities of causal association mining and anomaly identification among three types of data sources: text, images, and physiological signals. Its core structure may include a feature interaction layer, a causal inference layer, and an anomaly scoring layer, which can effectively and accurately locate the causal association between text tampering, image forgery, and physiological signal anomalies in the fused features. The threat representation vector refers to a fixed-dimensional feature vector with a preset format output after processing by the above interactive feature fusion algorithm. Each element in the vector corresponds to the feature representation of causal anomalies between different data sources, and may include core content such as anomaly type, anomaly confidence, the data source identifier associated with the anomaly, and the spatiotemporal information corresponding to the anomaly.

[0086] In the application, the fused features output above can be preprocessed, including feature normalization and screening of abnormal candidate regions. Then, the preprocessed fused features can be input into the feature interaction layer of the interactive feature fusion algorithm. Through dynamic interaction of the correlation features between text and image, image and physiological signals, and text and physiological signals, an interactive feature matrix is ​​generated. Next, the interactive feature matrix is ​​input into the causal inference layer to calculate the causal correlation weights between different feature dimensions, mining the causal logical relationships between text diagnostic conclusions and image lesions, image lesions and physiological signals, and text diagnosis and physiological signals. If a set of correlations violates normal medical logic, it is marked as a causal anomaly. Then, the anomaly scoring layer scores the confidence of the marked causal anomalies, filtering out anomaly features with confidence scores higher than a preset threshold. Finally, the anomaly features are encoded and integrated to generate a threat representation vector in a preset format. This threat representation vector will serve as the core trigger condition for subsequent risk assessment and security protection strategy adjustments, and will also be fed back to the federated learning framework to provide data support for the collaborative evolution of cross-institutional attack pattern libraries.

[0087] The data security protection method provided in the above embodiments completes the accurate feature extraction of text, images, and physiological signals through dedicated networks adapted to the characteristics of medical data (such as BioBERT, ResNet-50, and BiLSTM), ensuring the effectiveness of feature representation of multi-source heterogeneous data. It overcomes the spatiotemporal dimensional barriers of different types of medical data by leveraging a cross-modal attention mechanism, achieving efficient alignment and aggregation of multi-modal features. Through interactive feature fusion algorithms, it mines causal anomalies between data sources, improving the accuracy and interpretability of identifying security threats such as medical data tampering and image forgery. The final generated preset-format threat representation vector provides core data support for subsequent risk assessment and dynamic protection strategy adjustment, while also providing effective features for federated learning frameworks, facilitating the collaborative evolution of cross-institutional attack pattern libraries. Overall, it forms a complete technical chain of accurate feature extraction, efficient cross-modal fusion, accurate anomaly mining, and security protection support. While ensuring the effectiveness of medical data security protection, it also considers the professionalism and efficiency of medical business data processing, effectively solving the technical pain points of traditional medical data protection, such as difficulty in multi-source data fusion, low threat identification accuracy, and poor adaptability of protection strategies.

[0088] In some exemplary embodiments, such as Figure 3 As shown, S106 uses a large language model to perform semantic parsing of the operation instructions and jointly calculates the risk score based on the threat representation vector, including the following S302 to S306.

[0089] S302, acquire the operation command and the operation-related data corresponding to the operation command.

[0090] Operation instructions refer to the instructions issued by the operator (such as doctors, medical assistants, and system administrators) to the medical information system regarding access to and processing of medical data. These instructions include, but are not limited to, data retrieval, data modification, data download, and data sharing. Examples include retrieving chest CT images and corresponding medical records of patient Li XX from the past six months, or downloading a summary of the department's diagnostic and treatment data for diabetic patients from the past year. Operation-related data refers to auxiliary information directly associated with the operation instruction, used to justify the operation. This includes, but is not limited to, the operator's scope of responsibility (such as the authority boundaries of internists and radiology technicians), historical operation records (frequency of operations, data types, and time distribution within the past three months), relevance to the target patient (such as whether the operator is the patient's attending physician or involved in the patient's treatment team), and operation timestamp information (the time the operation was initiated, used to determine if it occurred outside of working hours).

[0091] In the application, the operation log collection module of the medical information system can capture operation instructions initiated by the operator, and synchronously record basic information such as the original text of the instruction, the initiating account, and the target data identifier. Subsequently, the system permission management module can call the operator's identity file to extract information such as the department, title, and list of responsibilities and permissions. The operation log database can be used to query the historical operation records of the initiating account in recent days, such as the last 30 days, and summarize and generate historical operation behavior records such as operation frequency statistics, operation data type distribution table, and historical operation time series. The patient diagnosis and treatment association database can be used to match the diagnosis and treatment association between the operator and the target patient to generate an association score. The precise timestamp of the operation is obtained through the system clock module and the corresponding time period attribute is marked (e.g., 8:00-18:00 on weekdays is the working period, and the rest is the non-working period). Finally, the above-mentioned operation instructions and various operation-related data are standardized in format and stored in a temporary data cache area for subsequent steps to call and input into the medical-specific large language model.

[0092] S304: Input the operation instructions and operation-related data into the large language model for semantic parsing to generate an operation intent vector.

[0093] The operation intent vector is a fixed-dimensional vector output after processing by the large language model. Vector elements accurately represent key attributes such as the core intent of the operation instruction (e.g., data retrieval, modification, batch download), the access object (e.g., specific patient, specific disease data), the operation scale (e.g., single data entry, batch data), and the urgency of the operation. In application, operation instructions and related data can be retrieved from a temporary data cache, then input into the large language model to output the operation intent vector. A preliminary rationality assessment of the operation behavior can also be output simultaneously. Finally, the operation intent vector and the preliminary rationality assessment result can be stored in a feature cache for subsequent steps in S306.

[0094] S306: Risk score is obtained by jointly calculating the operational intent vector and threat representation vector.

[0095] The joint computation can employ a combination algorithm of attention-weighted fusion and multi-dimensional weighted summation. In application, the operation intent vector generated by S304 and the threat representation vector generated by S210 can be retrieved from the feature cache. These two types of vectors are then input into the attention-weighted fusion module to calculate their correlation weights, uncovering the matching relationship between the operation intent and data anomalies (e.g., when the operation intent is to batch download medical records and the threat representation vector indicates abnormal text data tampering, the correlation weight is increased). Next, based on the risk assessment dimensions, scores for each dimension can be calculated using the correlation weights, and the final risk score can be calculated using a weighted summation formula. Finally, the risk score and detailed scores for each dimension are stored in the risk assessment database, serving as the core decision-making basis for subsequent upgrades to encryption levels and triggering enhanced authentication processes.

[0096] The data security protection method provided in the above embodiments acquires operation instructions and corresponding operation-related data, providing complete and accurate basic input for subsequent risk assessment and avoiding assessment bias caused by missing data. Then, it utilizes a large language model to perform deep semantic analysis on the operation instructions and operation-related data, transforming abstract operation intentions into quantifiable and computable operation intention vectors. This effectively solves the technical challenge of traditional methods failing to accurately capture the core intentions behind complex operations, improving the accuracy and intelligence level of operation intention recognition. Finally, a risk score is obtained through the joint calculation of the operation intention vector and threat representation vector, achieving a deep correlation analysis between operation intentions and potential threats. This overcomes the limitations of risk assessment based solely on single-dimensional data, providing a more comprehensive and objective reflection of the risk level of operational behavior, offering a reliable basis for the precise formulation of subsequent risk prevention and control measures, and significantly improving the targeting and effectiveness of risk identification.

[0097] In some exemplary embodiments, the data security protection method further includes: when the risk score exceeds a preset first threshold, inputting the operation instructions and threat characterization vector into a large language model and outputting a causal correlation audit report; and updating the multimodal fusion network based on the audit results corresponding to the causal correlation audit report.

[0098] The causal correlation audit report refers to a structured report generated by integrating multi-source data through a large language model based on the security audit needs of medical scenarios. It clearly presents the causal logic between operational behavior and risk characteristics. Its content includes, but is not limited to, basic operational information (operator, operation time, operation content, etc.), risk scoring and triggering basis, operational intent tracing results, abnormal feature correlation links, audit conclusions, and data access tracing, providing intuitive and interpretable evidence for security handling. The audit result refers to the conclusive opinion formed after verifying and confirming the causal correlation audit report. A dual confirmation mode of manual review and automatic system verification can be adopted. Manual review focuses on verifying key information such as the authenticity of the operation and the legality of the account, while automatic system verification confirms the novelty of abnormal features by comparing with a historical risk case database. Finally, a clear result is output: audit passed (risk / abnormality confirmed) or audit failed (misjudgment excluded).

[0099] The data security protection method provided in the above embodiments, through the causal correlation audit report generated by the large language model, can clearly present the intent, chain, and abnormal characteristic correlation logic of the violation operation, providing security auditors with intuitive and interpretable audit basis, helping to quickly handle high-risk events, and effectively solving the problems of vague risk tracing and unclear causal logic in traditional audits. At the same time, updating the multimodal fusion model based on the audit results can enable the model to accurately learn and identify violation characteristics, significantly improving the system's ability to identify unknown threats. It successfully overcomes the technical bottleneck that traditional static protection models are difficult to adapt to dynamic threats to medical data security. While ensuring the protection strength in high-risk scenarios, it avoids interference with normal medical business by indiscriminate protection, achieving a dynamic balance between security protection and business efficiency. Moreover, the structured audit report completely retains the operation trajectory and risk judgment basis, which meets the compliance requirements of medical data security audit and provides strong support for liability determination.

[0100] In some exemplary embodiments, such as Figure 4 As shown, the data security protection method also includes the following S402 to S408.

[0101] S402, Obtain threat feature information generated locally by the medical institution based on threat representation vector.

[0102] Threat feature information is a set of core information obtained after feature extraction and filtering of threat representation vectors. It can accurately characterize the key attributes of potential threats in medical data. In some exemplary embodiments, threat feature information includes a threat feature vector summary and / or parameter update gradient. The threat feature vector summary is a condensed representation of the threat representation vector, retaining the core semantics and feature dimensions of threat identification, and can be quickly used for cross-institutional data interaction and comparison. The parameter update gradient consists of model optimization parameters generated during the iterative training of the local threat detection model based on the threat representation vector, reflecting the adjustment requirements of the latest local threat patterns on the model.

[0103] Medical institutions can achieve this locally through a pre-defined local feature processing module. Specifically, the local system can call a data interface to obtain the generated threat representation vector from the threat perception layer; subsequently, the local feature processing module can activate a feature filtering mechanism to filter out redundant and invalid feature dimensions based on a pre-defined medical security threat feature dictionary; next, a threat feature vector summary can be generated through a feature condensation algorithm, while extracting the parameter update gradients of the local threat detection model during training for the threat representation vector, and finally integrating them to form complete threat feature information and storing it in a local security database, awaiting subsequent privacy protection processing.

[0104] S404 performs privacy protection processing on threat signature information to generate privacy-protected data.

[0105] Privacy-preserving processing refers to the process of masking the associated information of original medical data through encryption, perturbation, and other technical means while retaining the core identification value of threat feature information. In some exemplary embodiments, differential privacy technology can be used to achieve this. The core purpose is to prevent the leakage of patient personal privacy and internal medical institution data details during cross-institutional data interaction, ensuring compliance with relevant regulatory requirements. Privacy-preserving data is usable data formed after threat feature information has undergone privacy-preserving processing. It retains the core feature dimensions of the threat pattern, supporting the aggregation training of cloud-based federated learning, while eliminating the direct association with the original medical data to avoid the risk of privacy leakage. It may include a threat feature vector summary and / or parameter update gradient after adaptive noise addition processing.

[0106] In applications, for threat feature information associated with highly sensitive data such as patient gene testing results and infectious disease diagnosis records, a high-intensity adaptive noise-adding strategy can be adopted to further enhance privacy protection by increasing the noise perturbation amplitude; for threat feature information associated with ordinary outpatient medical records, the noise intensity is appropriately reduced to ensure privacy and security while also taking into account the model training accuracy.

[0107] S406 aggregates privacy-protected data sent by various medical institutions through a cloud-based federated learning architecture to train and generate a global threat detection model.

[0108] The cloud-based federated learning architecture is a distributed collaborative training system deployed in the cloud. It consists of a federated coordination server, a global model training unit, a secure communication gateway, and a data format verification module. It enables collaborative utilization of cross-institutional data and model co-construction without obtaining the original medical data from each institution, serving as a core carrier supporting cross-institutional threat protection collaboration. Privacy-preserving data aggregation refers to the process by which the cloud architecture uniformly integrates, cleans, and fuses privacy-preserving data uploaded by various institutions. For example, an improved FedProx aggregation algorithm can be used. First, the data format verification module removes privacy-preserving data that does not meet the specifications. Then, gradient alignment and weight fusion are performed on the valid data to uncover common patterns and specific characteristics of threat features from different institutions, avoiding the problem of insufficient model generalization ability caused by insufficient data samples from a single institution. The global threat detection model is a general threat identification model trained on aggregated privacy-preserving data. It has the ability to detect threats across institutions and multiple scenarios. It can identify typical threats common to different medical institutions as well as new threats in specific scenarios. Its model structure may include a feature input layer, a cross-institutional feature fusion layer, a threat classification layer and an output layer. It can output key identification results such as threat type, threat level and confidence level.

[0109] S408 performs lightweight processing on the global threat detection model to obtain a lightweight model, and then deploys the lightweight model to the local systems and edge devices of various medical institutions.

[0110] Lightweight processing refers to the process of reducing the number of model parameters and computational complexity by using model compression techniques while retaining the core threat identification capabilities of the global threat detection model. For example, knowledge distillation can be used. For instance, a global ResNet-101 architecture model trained in the cloud can be distilled into a lightweight ResNet-18 architecture model, reducing computational overhead while maintaining threat identification accuracy. The lightweight model, obtained after lightweight processing, is a threat detection model adapted for terminal deployment. It features small size, fast running speed, and low resource consumption, enabling stable operation on local systems in medical institutions and edge devices with limited computing power. Simultaneously, it fully retains the cross-institutional threat identification capabilities of the global model, allowing real-time threat detection of locally collected medical data and outputting accurate threat judgment results.

[0111] During deployment, the federated coordination server of the cloud-based federated learning architecture first performs compatibility verification on the lightweight model to ensure it can adapt to the local system environments of different medical institutions and the hardware configurations of various edge devices. After verification, the lightweight model is distributed to each medical institution through a secure communication link. During distribution, the model file is encrypted and protected using national cryptographic algorithms to prevent the model from being tampered with or stolen. After receiving the model, each medical institution completes the installation and debugging of the lightweight model on its local server and edge devices through the local deployment module. After debugging, the model is started in real-time running mode to achieve real-time threat detection of local medical data on the edge.

[0112] The data security protection method provided in the above embodiments, by generating threat feature information locally and processing it for privacy protection, not only safeguards patient privacy and medical institution data security, and complies with relevant medical data regulations, but also retains the core identification value of threat features, resolving the contradiction between privacy and usability in traditional cross-institutional data sharing. Furthermore, by leveraging the aggregation training of a cloud-based federated learning architecture, it integrates threat data resources from multiple institutions, effectively compensating for the poor generalization ability of models caused by insufficient samples from a single medical institution. This enables the generated global threat detection model to have threat identification capabilities across institutions and multiple scenarios, improving the accuracy of responding to new and complex threats. In addition, through lightweight processing and edge deployment, it adapts to the differences in system resources of medical institutions of different sizes, achieving real-time edge response for threat detection, reducing cloud dependence and data transmission overhead. At the same time, it forms a closed-loop mechanism of local feature extraction, cloud collaborative training, and edge deployment iteration, ensuring that the threat detection capabilities of each medical institution are continuously upgraded with the optimization of the global model, comprehensively improving the synergy, adaptability, and practical effectiveness of the medical data security protection system.

[0113] In some exemplary embodiments, the threat feature information includes a threat feature vector summary and / or parameter update gradient. Specifically, step S404 involves performing privacy-preserving processing on the threat feature information to generate privacy-preserving data, including: adaptively adding noise to the threat feature vector summary and / or parameter update gradient locally generated by the medical institution using differential privacy technology to generate privacy-preserving data. For details, please refer to the foregoing description, which will not be repeated here.

[0114] In some exemplary embodiments, data security protection further includes: performing anomaly detection on image data and text data in heterogeneous data respectively, generating corresponding image anomaly indicators and text anomaly indicators; determining whether there is an adversarial attack based on the combined judgment result of the image anomaly indicators and text anomaly indicators; triggering a security isolation mechanism and tamper-proofly storing the attack evidence if the judgment result is yes; and sending the feature information of the adversarial attack to the cloud federated learning architecture to update the global threat detection model.

[0115] In applications, image data can be analyzed in the frequency domain and its texture consistency verified to generate image anomaly scores, i.e., anomaly indicators. Similarly, text data can be analyzed for causal inconsistencies to generate text anomaly scores, i.e., text anomaly indicators. For example, high-frequency energy analysis in the frequency domain can be performed on image data to identify anomalous noise distributions in GAN-generated images. Simultaneously, texture feature analysis using the gray-level co-occurrence matrix can capture microscopic inconsistencies in forged regions, comprehensively generating an image anomaly score. For example, the Causal-BERT model can be used to detect causal inconsistencies in text data, tracing medical record modifications and analyzing the causal relationships between diagnostic conclusions and laboratory test indicators and physiological signal data to generate text anomaly scores.

[0116] If the image anomaly score exceeds a preset second threshold and the text anomaly score exceeds a preset third threshold, an adversarial attack is determined to exist. If an adversarial attack is determined, a hardware-level trusted execution environment (TEE) is immediately triggered to isolate the suspicious data stream and initiate a blockchain evidence storage process. Attack evidence (including but not limited to attack time, operating account, and tampered content) is recorded and stored immutably. Furthermore, the dynamic protection engine can adjust the permission policies of the corresponding access accounts. Simultaneously, adversarial attack feature information (including but not limited to frequency domain anomaly patterns, texture inconsistencies, and causal contradictions encoded into standardized feature vectors) is extracted and uploaded to a cloud-based federated learning framework. The cloud-based federated learning server uses this adversarial attack feature information as incremental training samples to update the global threat detection model and distributes the updated lightweight model to various medical institutions. For example, the cloud-based federated learning server can use standardized feature vectors as incremental training samples, update the global threat detection model through an asynchronous aggregation algorithm, compress it into a lightweight model through knowledge distillation, and then distribute it to the local systems and edge devices of various medical institutions, achieving co-evolution of protection strategies.

[0117] The data security protection method provided in the above embodiments achieves multi-dimensional and accurate perception of adversarial attacks by separately performing anomaly detection on image and text data in heterogeneous data and generating corresponding anomaly indicators. This effectively overcomes the limitations of traditional single-data type detection, which is prone to false positives and false negatives. In particular, it can accurately identify new adversarial attack behaviors such as generative AI-forged medical images and malicious tampering with electronic medical records. Furthermore, based on the combined judgment mechanism of dual-dimensional anomaly indicators, the causal relationship between the two types of data can be combined to improve the accuracy of adversarial attack judgment, avoiding misjudgments caused by a single abnormal indicator. This ensures strong security protection while reducing interference with normal medical operations. Furthermore, upon determining the existence of adversarial attacks, the security isolation mechanism can quickly block the spread of threats and prevent attack payloads from penetrating the core areas of the medical information system. The tamper-proof evidence storage provides a reliable basis for attack tracing and liability determination, meeting the compliance requirements of medical data security audits. In addition, the adversarial attack characteristic information is fed back to the cloud-based federated learning architecture to update the global threat detection model, forming a closed-loop optimization mechanism of detection, defense, and model iteration. This enables the global model to continuously adapt to new adversarial attack scenarios, simultaneously improving the adversarial attack defense capabilities of various medical institutions, and comprehensively strengthening the proactive defense level and continuous evolution capability of the medical data security protection system.

[0118] In some exemplary embodiments, a data security protection method is provided, which can be applied to medical information systems. In this application, a multimodal perception module can be deployed at the data acquisition end to simultaneously collect multi-source data such as electronic medical record text, medical images, and physiological signals, and transmit them to a threat perception center via a dedicated network. Multimodal feature extraction algorithms are used to preprocess various types of data, providing standardized input for subsequent threat analysis. Specifically, text data undergoes semantic parsing and keyword extraction using natural language processing technology; key region features are extracted from medical images using convolutional neural networks; and dynamic change patterns of physiological signals are captured using a time-series analysis model. Then, the extracted features are input into a multimodal fusion model, which aligns the spatiotemporal features of different data sources through a cross-modal attention mechanism to generate a unified threat representation vector. The threat perception center assesses the risk level of data flow in real time based on the threat vector. When abnormal access patterns or potential attack behaviors are detected, a blocking mechanism is immediately triggered, setting dynamic access control points on the data transmission path, cutting off suspicious connections, and marking and tracking related data flows. Simultaneously, threat event details are fed back to the dynamic protection engine, providing a basis for subsequent security strategy adjustments.

[0119] Then, a medical-specific large language model is integrated into the medical information system to perform semantic understanding and intent parsing of user operation commands, and to construct a dynamic trust assessment model based on user historical behavior. Access permissions and encryption strategies are adjusted in real time based on the assessment results to achieve precise security control. Specifically, when a user initiates a data access request, the large language model performs contextual analysis of the operation command, considering multiple dimensions such as user role, scope of responsibility, operation time, and data sensitivity, to calculate a risk score for the operation behavior. If the risk score exceeds a preset threshold, the system automatically triggers an enhanced authentication process, requiring the user to provide personalized biometric verification, such as fingerprint, iris, or facial recognition. Simultaneously, the encryption algorithm and key length are dynamically adjusted according to the data sensitivity and operation scenario. High-sensitivity data uses national cryptographic algorithms for high-strength encryption, while low-sensitivity data uses lightweight encryption schemes to ensure business efficiency. The dynamic protection engine continuously monitors user operation behavior. When abnormal operation patterns or signs of permission abuse are detected, access control policies are updated in real time, restricting user access to specific data and auditing related operations to ensure the compliance and security of data access.

[0120] Furthermore, each medical unit deploys a lightweight local model, performing feature extraction and preliminary analysis of medical data only within the institution. Raw data is strictly retained locally to ensure data privacy. Threat feature summaries or encrypted parameter update gradients are uploaded to the federated learning server via a secure communication channel, enabling cross-institutional knowledge sharing and collaborative model optimization. Specifically, the local model continuously learns new attack features within the institution, periodically extracts and encrypts threat feature summaries, and transmits them to the federated server via a secure protocol. The federated server uses advanced aggregation algorithms to integrate feature summaries and parameter updates from various institutions, generating a global threat detection model. It then uses knowledge distillation technology to compress the model into a miniaturized version suitable for edge device deployment. The distilled model is then distributed to edge devices in each medical institution, such as CT scanners and monitors, enabling real-time threat detection and local protection strategy updates. Simultaneously, the federated learning framework dynamically adjusts the weight allocation of each institution in model training based on its data quality, sample richness, and contribution to the global model. This incentivizes medical institutions to provide high-quality data and abundant attack samples, creating a positive cycle that continuously improves the threat detection capabilities and adaptability of the global model.

[0121] To address generative AI forgery threats, a multimodal threat analysis module is constructed, integrating frequency domain analysis and causal reasoning technologies for deep detection of medical images and text medical records. A hardware-level trusted execution environment (HEX) is used to achieve threat isolation and rapid response, ensuring the system's secure and stable operation. Specifically, in the image domain, frequency domain analysis identifies abnormal noise patterns and texture inconsistencies in medical images, detecting forged images generated by GANs; in the text domain, a causal reasoning model traces medical record modifications, analyzing causal contradictions between diagnostic conclusions and examination indicators to determine the likelihood of record tampering. When suspicious data is detected, the system immediately isolates the relevant data stream to the HEX, leveraging its isolation and protection features to prevent the threat from spreading throughout the system. Simultaneously, a blockchain evidence storage process is initiated to record and store attack evidence immutably, providing strong support for subsequent attack tracing and accountability. A dynamic protection engine adjusts access permissions for relevant accounts in real time based on threat detection results and promptly feeds new attack characteristics back to the federated learning framework, achieving collaborative evolution of protection strategies and enhancing the overall system's protection capabilities.

[0122] In practical applications and technological evolution, this application can achieve functional optimization and scenario adaptation through various variations. On one hand, in the multimodal threat perception and blocking module, an incremental learning mechanism can be introduced, enabling the model to continuously learn new attack characteristics, dynamically update the threat detection model, and improve the ability to identify unknown threats. Simultaneously, edge computing technology can be combined to push some threat analysis functions to edge devices at the data acquisition end, reducing data transmission latency and improving threat response speed, especially suitable for medical scenarios with high real-time requirements. On the other hand, for the dynamic access control and encryption policy adjustment module driven by a large language model, the introduction of multi-factor authentication and risk-based adaptive authentication mechanisms can be explored to further enhance the security and flexibility of user authentication. Furthermore, new encryption algorithms and key management schemes can be researched and developed to adapt to constantly changing security needs and technological challenges. Under the federated learning framework, support for more types of medical institutions and data sources to participate in co-evolution can be expanded, improving cross-institutional collaboration efficiency and data transmission security by optimizing communication protocols and data encryption schemes. At the same time, in-depth research can be conducted on model compression and optimization technologies to enable the global model to better adapt to resource-constrained edge devices, improving overall protection performance. For the collaborative defense module of multimodal threat analysis and hardware-level trusted execution environment, we can explore integrating more multi-dimensional threat detection technologies, such as anomaly detection based on behavioral analysis and malicious traffic identification based on artificial intelligence, to further enhance the detection capability against complex attacks. In addition, we can research enhancement schemes for the hardware trusted execution environment to improve its ability to combat advanced persistent threats and ensure the security and reliability of critical system data and operations.

[0123] This application addresses technical bottlenecks in the field of medical data security, such as fragmented protection of multi-source heterogeneous data, lagging dynamic threat response, and an imbalance between privacy protection and data availability. It proposes a dynamic security protection system and method for medical data based on a multimodal large language model (MLLM). By constructing a cross-modal threat representation learning framework, it integrates multi-dimensional data such as electronic medical records, medical images, and physiological signals, overcoming the spatiotemporal discontinuity problem of traditional single-modal detection and achieving a leap from isolated analysis to unified understanding of threat perception. Combining the contextual parsing capabilities of a medical-specific large language model (LLM), it dynamically assesses the intent of medical operations and adjusts encryption strength and access permission policies in real time while ensuring diagnostic and treatment efficiency—for example, automatically triggering biometric secondary authentication and switching to national cryptographic algorithms when a doctor accesses highly sensitive images. Addressing the challenges of data silos and privacy barriers between medical institutions, a lightweight federated learning framework is designed, allowing hospitals to collaboratively evolve threat detection models without sharing raw data. Differential privacy and knowledge distillation techniques are used to balance the accuracy of threat intelligence sharing with the needs of patient privacy protection. In addition, the system integrates an anti-attack defense module, which uses frequency domain analysis and causal reasoning models to simultaneously detect image tampering and text forgery, and uses a hardware-level trusted execution environment (TEE) to achieve attack isolation, forming a proactive protection system covering the entire lifecycle of data collection, transmission, and storage.

[0124] This application, through the aforementioned technical solutions, significantly enhances the intelligence and adaptability of the medical data security protection system. Multimodal threat characterization technology overcomes the limitations of traditional single-modal detection, significantly enhancing the ability to identify lateral penetration attacks and multi-stage composite threats. For example, by associating textual medical record tampering with image forgery, it achieves accurate tracing of covert attack chains. The dynamic protection mechanism driven by a large language model abandons static rule dependence, enabling access control and encryption strategies to be flexibly adjusted according to operational scenarios, ensuring the availability of clinical data while effectively curbing unauthorized access and account theft risks. The application of the federated learning framework significantly optimizes the efficiency of cross-institutional threat intelligence sharing, shortens attack pattern update delays, and reduces communication overhead through differential privacy and model compression technologies. The system-integrated anti-adversarial attack module detects generative AI-forged medical image and medical record tampering behaviors in real time, and, combined with a hardware-level trusted execution environment, achieves rapid isolation and rollback of threat data.

[0125] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0126] Based on the same inventive concept, this application also provides a data security protection system for implementing the data security protection method described above. The solution provided by this device is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more data security protection device embodiments provided below can be found in the limitations of the data security protection method described above, and will not be repeated here.

[0127] In some exemplary embodiments, such as Figure 5 As shown, a data security protection system 500 is provided, applied to a medical information system. The data security protection system 500 includes: a multimodal perception module 501, used to acquire heterogeneous data in the medical information system; the heterogeneous data includes text data, image data, and physiological signal data; a threat representation learning module 502, used to input the heterogeneous data into a multimodal fusion model and output a threat representation vector in a preset format; the threat representation vector is used to represent the degree of anomaly of the heterogeneous data; and a dynamic protection module 503, used to perform semantic parsing of operation instructions using a large language model and jointly calculate a risk score based on the threat representation vector; the risk score is used to represent the risk level of the operation instruction initiating access to data in the medical information system; if the risk score exceeds a preset first threshold, the encryption level of the accessed data is upgraded and an enhanced authentication process is triggered. The threat representation learning module 502 can be understood as the aforementioned threat perception center, and the dynamic protection module 503 can be understood as the aforementioned dynamic protection engine; for details, please refer to the aforementioned related descriptions, which will not be repeated here.

[0128] Please continue reading. Figure 5In some exemplary embodiments, the data security protection system 500 further includes a federated learning module 504, used to aggregate privacy-protected data sent by various medical institutions, train and generate a global threat detection model, perform lightweight processing on the global threat detection model to obtain a lightweight model, and deploy the lightweight model to the local systems and edge devices of each medical institution; wherein, the privacy-protected data is threat feature information generated locally by the medical institution based on threat representation vectors, and the threat feature information is generated by performing privacy-protected processing. The federated learning module 504 can be understood as the aforementioned federated learning framework deployed in the cloud, as detailed in the foregoing related introduction, and will not be repeated here.

[0129] Please continue reading. Figure 5 In some exemplary embodiments, the data security protection system 500 further includes: an adversarial attack defense module 505, used to perform anomaly detection on image data and text data in heterogeneous data respectively, and generate corresponding image anomaly indicators and text anomaly indicators; determine whether there is an adversarial attack based on the combined judgment result of the image anomaly indicators and text anomaly indicators; if the judgment result is yes, trigger a security isolation mechanism and perform tamper-proof evidence storage of the attack; and send the feature information of the adversarial attack to the cloud federated learning architecture to update the global threat detection model.

[0130] It is understandable that with the widespread deployment of medical IoT devices, the system architecture can be further optimized to strengthen edge computing and cloud collaboration. At the data acquisition end, adding an edge preprocessing module enables terminals such as smart monitors and portable ultrasound devices to possess preliminary threat perception capabilities. For example, lightweight convolutional neural networks can be used to detect the integrity of medical images in real time at the device end, uploading only suspected tampered image segments to the cloud for analysis, significantly reducing data transmission bandwidth requirements. The cloud focuses on cross-institutional threat intelligence integration and global model training, dynamically allocating computing resources through task scheduling algorithms to prioritize high-risk data samples. Simultaneously, a dual trusted execution environment at both the edge and cloud levels is designed to ensure the confidentiality and integrity of data during transmission and processing, adapting to the architectural needs of future explosive growth in medical data volume.

[0131] Furthermore, it enables the fusion of proactive immunity and defense-in-depth for security protection. This means shifting from passive defense to proactive immunity, endowing the system with self-healing capabilities. Building upon existing dynamic protection, an immune algorithm module is integrated to preset a health baseline for medical data security. When abnormal behavior deviates from the baseline, not only is a blocking mechanism triggered, but a data repair process is also automatically initiated. For example, for tampered electronic medical records, erroneous diagnostic conclusions are automatically corrected based on historical data versions and medical knowledge graph verification. Simultaneously, a defense-in-depth system is constructed, deploying intelligent firewalls at the network boundary and using deep packet inspection technology to identify potential threats in medical-specific protocols. At the application layer, identity spoofing protection is strengthened, and blockchain technology is used to record the entire user operation process, ensuring that every access behavior is traceable and auditable, forming a comprehensive, multi-layered security protection ecosystem.

[0132] Furthermore, it allows for continuous evolution of intelligence and enhanced human-machine collaboration. With a focus on the continuous evolution of artificial intelligence technology, it reserves model upgrade interfaces to adapt to more powerful multimodal large language models in the future. For example, when the next-generation medical-specific LLM can understand the three-dimensional structure and pathological features of medical images, the intent parsing accuracy of the dynamic protection engine can be optimized, enabling fine-grained control over data access in complex surgical planning scenarios. Simultaneously, it strengthens the human-machine collaboration mechanism, developing an interactive threat analysis interface that allows security analysts to intuitively understand the hidden attack chains detected by the system and adjust protection strategies through natural language feedback. For example, analysts can quickly authorize the isolation of suspicious data via voice commands or request the system to generate simulation reports for specific attack scenarios, improving the overall adaptability and flexibility of the protection.

[0133] Furthermore, incremental learning and lightweight modifications for multimodal threat perception are possible. Addressing the resource constraints of medical IoT devices, the multimodal threat perception module undergoes a lightweight modification. Model distillation technology is employed to compress complex multimodal fusion models into miniaturized versions suitable for edge device deployment. For example, the ResNet network is distilled into a lighter architecture, reducing the number of parameters while retaining the ability to extract key medical image features. An incremental learning mechanism is introduced, enabling edge devices to continuously learn new threat features without transmitting data back to the source. For instance, when a new image forgery pattern is detected, only the parameters of the last few convolutional layers of the edge model are updated, ensuring the device adapts to new attack methods in real time while protecting patient privacy.

[0134] Furthermore, it achieves asynchronous updates and differential privacy enhancements in federated learning. The federated learning framework is optimized, employing an asynchronous update mechanism to improve cross-institutional collaboration efficiency. Medical institutions are allowed to flexibly upload parameter updates based on their own data processing progress. The cloud server integrates asynchronously arriving update information through a conflict detection algorithm, avoiding the long-tail latency problem caused by traditional synchronous updates. The differential privacy protection level is enhanced by introducing an adaptive noise injection algorithm. The noise intensity is dynamically adjusted according to the sensitivity of the uploaded data. For example, for highly sensitive data features containing patient genetic information, the noise perturbation amplitude is increased to ensure that the risk of privacy leakage is reduced to below the medical industry's security standards, thereby increasing the enthusiasm of medical institutions to participate in federated learning and improving the quality of data sharing.

[0135] Furthermore, it enables enhanced cross-modal collaborative defense against adversarial attacks. The detection dimensions of the adversarial attack module are expanded, adding behavioral analysis and traffic feature detection. In addition to existing frequency domain analysis and causal reasoning techniques, user operation behavior models are integrated to identify abnormal data access patterns. For example, when an account frequently accesses image data unrelated to its own department within a short period, accompanied by abnormal peaks in network traffic, it is identified as potential data theft. By combining cross-modal information correlation, such as comparing the geographical location of the operating terminal with doctor scheduling information to achieve multi-dimensional threat assessment, it enhances the isolation capabilities of the hardware-level trusted execution environment. For advanced persistent threats (APTs), fine-grained resource access control is designed to ensure that attack payloads cannot move laterally to the core data storage area.

[0136] Each module in the aforementioned data security protection system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0137] In some exemplary embodiments, a medical data security protection system is provided, including a medical information system and the data security protection system of any of the foregoing embodiments.

[0138] In some exemplary embodiments, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6As shown. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used to store data in a medical information system. The network interface of the computer device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a data security protection method. Those skilled in the art will understand that... Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0139] In some exemplary embodiments, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the aforementioned method.

[0140] In some exemplary embodiments, a computer-readable storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned method.

[0141] In some exemplary embodiments, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the aforementioned method.

[0142] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0143] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0144] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A data security protection method, characterized in that, Applied to a medical information system, the method includes: Acquire heterogeneous data from the medical information system; the heterogeneous data includes text data, image data, and physiological signal data; The heterogeneous data is input into a multimodal fusion model, which outputs a threat characterization vector in a preset format; the threat characterization vector is used to represent the degree of anomaly of the heterogeneous data. A large language model is used to perform semantic parsing on the operation instructions, and a risk score is jointly calculated based on the threat representation vector; the risk score is used to represent the degree of risk of the operation instructions initiating access to data in the medical information system. If the risk score exceeds a preset first threshold, the encryption level of the accessed data is upgraded and an enhanced authentication process is triggered.

2. The method according to claim 1, characterized in that, The multimodal fusion model includes a natural language processing network, a convolutional neural network, a temporal neural network, a cross-modal attention mechanism, and an interactive feature fusion algorithm; the step of inputting the heterogeneous data into the multimodal fusion model and outputting a threat representation vector in a preset format includes: The text data is input into the natural language processing network for feature extraction to generate text features; The image data is input into the convolutional neural network for feature extraction to generate image features; The physiological signal data is input into the temporal neural network for feature extraction to generate physiological features; The cross-modal attention mechanism is used to perform spatiotemporal alignment of the text features, image features, and physiological features to generate fused features; The interactive feature fusion algorithm is used to mine causal anomalies between different data sources in the fused features and generate a threat characterization vector in a preset format.

3. The method according to claim 1, characterized in that, The step of using a large language model to perform semantic parsing of operation instructions and jointly calculating a risk score based on the threat representation vector includes: Acquire the operation instruction and the operation-related data corresponding to the operation instruction; The operation instructions and operation-related data are input into the large language model for semantic parsing to generate an operation intent vector. The risk score is obtained by jointly calculating the operational intent vector and the threat representation vector.

4. The method according to claim 1, characterized in that, The method further includes: If the risk score exceeds a preset first threshold, the operation instructions and the threat representation vector are input into the large language model, and a causal correlation audit report is output. The multimodal fusion network is updated based on the audit results corresponding to the causal correlation audit report.

5. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain threat feature information generated locally by the medical institution based on the threat representation vector; The threat signature information is processed for privacy protection to generate privacy-protected data; By using a cloud-based federated learning architecture, the privacy-protected data sent by various medical institutions is aggregated and trained to generate a global threat detection model. The global threat detection model is lightweighted to obtain a lightweight model, which is then deployed to the local systems and edge devices of various medical institutions.

6. The method according to claim 5, characterized in that, The threat feature information includes a threat feature vector summary and / or parameter update gradients; wherein... The step of performing privacy protection processing on the threat feature information to generate privacy-protected data includes: Differential privacy technology is used to add noise to the threat feature vector summary and / or parameter update gradient of the medical institution's local data to generate the privacy-protected data.

7. The method according to any one of claims 1-4, characterized in that, The method further includes: Anomaly detection is performed on the image data and text data in the heterogeneous data respectively to generate corresponding image anomaly indicators and text anomaly indicators. Based on the combined judgment results of the image anomaly indicators and text anomaly indicators, a judgment result on whether there is an adversarial attack is determined. If the judgment result is yes, a security isolation mechanism is triggered and the attack evidence is preserved in an immutable manner; The characteristic information of the adversarial attack is sent to the cloud-based federated learning architecture to update the global threat detection model.

8. A data security protection system, characterized in that, Applied to medical information systems, including: A multimodal perception module is used to acquire heterogeneous data in the medical information system; the heterogeneous data includes text data, image data, and physiological signal data. The threat representation learning module is used to input the heterogeneous data into a multimodal fusion model and output a threat representation vector in a preset format; the threat representation vector is used to represent the degree of anomaly of the heterogeneous data. The dynamic protection module is used to perform semantic parsing of operation instructions using a large language model and jointly calculate a risk score based on the threat representation vector. The risk score is used to represent the risk level of the operation instruction initiating access to data in the medical information system. If the risk score exceeds a preset first threshold, the encryption level of the accessed data is upgraded and an enhanced authentication process is triggered.

9. The system according to claim 7, characterized in that, The system also includes: The federated learning module is used to aggregate privacy-protected data sent by various medical institutions, train a global threat detection model, perform lightweight processing on the global threat detection model to obtain a lightweight model, and deploy the lightweight model to the local systems and edge devices of various medical institutions. The privacy-protected data is threat feature information generated locally by the medical institutions based on the threat representation vector, and the threat feature information is generated after privacy protection processing.

10. The system according to claim 7, characterized in that, The system also includes: The adversarial attack defense module is used to perform anomaly detection on the image data and text data in the heterogeneous data, respectively, and generate corresponding image anomaly indicators and text anomaly indicators; based on the combined judgment result of the image anomaly indicators and text anomaly indicators, it determines whether an adversarial attack exists; if the judgment result is yes, it triggers a security isolation mechanism and performs tamper-proof evidence storage of the attack; and sends the feature information of the adversarial attack to the cloud federated learning architecture to update the global threat detection model.

Images