Method for preventing leakage of sovereign data in combination with multi-mode deception feature perception

Abstract

The invention provides a sovereign data leakage prevention method combined with multi-mode deception feature perception, and relates to the technical field of computer security, the current feature is input to a corresponding confidence model to calculate the confidence corresponding to the current feature, and the current feature state is judged according to the confidence corresponding to the current feature and a multi-level dynamic threshold; and constructing a dynamic deception feature database based on the normal feature tag, the doubt feature tag and the deception feature tag, and inputting the current feature into the dynamic deception feature database for deception feature perception. By constructing a multistage dynamic threshold mechanism based on historical sample distribution and combining a supervised classification model, a self-encoding model and a time sequence discrimination model, fine discrimination of normal, doubt and abnormal states is realized, and a multi-modal cross validation mechanism and a feedback re-determination process are introduced to form a closed-loop discrimination system from perception, verification to recheck. And the misjudgment rate caused by boundary samples and system noise is reduced.

Description

Technical Field

[0001] The present application relates to the field of computer security technology, and in particular to a sovereign data leakage prevention method combined with multimodal deception feature perception. Background Art

[0002] As the role of data resources in areas such as sovereign security, corporate compliance and intelligent decision-making continues to deepen, data leakage prevention technology has gradually evolved from static rule matching to behavioral pattern perception. Traditional methods are mostly based on single modal features for judgment, mainly through keyword filtering, traffic detection or behavioral log rules to identify and control sensitive data.

[0003] However, current data leakage behaviors are becoming increasingly covert, especially in complex environments where multimodal data are cross-combined. Attackers circumvent traditional protection models through inter-modal feature disguise, deceptive perturbations, and cross-modal migration, which seriously threatens the integrity and credibility of sovereign data assets.

[0004] Existing multimodal data leakage prevention systems generally have the following problems: First, it lacks the ability to accurately distinguish between "abnormal features" and "deceptive features", making it easy to misjudge system noise or boundary samples as risky data, resulting in a high false alarm rate; Second, most of them rely on static threshold settings and lack a dynamic discrimination mechanism based on sample history confidence, making it difficult to adapt to the distribution differences between different modalities; Third, the review process for data that cannot distinguish abnormal or normal data lacks hierarchical management and feedback loops, making it impossible to re-identify and iteratively defend against potential fraudulent behaviors, and making it difficult to support the identification and optimization of new fraud features. Summary of the Invention

[0005] This application provides a sovereign data leakage prevention method that combines multimodal deception feature perception, aiming to solve the problems of poor recognition accuracy, slow response speed and non-closed feedback loop when processing complex deception behaviors in multimodal data environments in related technologies.

[0006] To achieve the above objectives, this application provides a sovereign data leakage prevention method that combines multimodal deception feature perception, including the following: Extract each modal feature based on the acquired multimodal data.

[0007] A confidence model corresponding to each modal feature is constructed, and a multi-level dynamic threshold is set for each modal feature; the multi-level dynamic threshold includes a normal dynamic threshold, a questionable dynamic threshold, and an abnormal dynamic threshold.

[0008] The current feature is input into the corresponding confidence model to calculate the confidence corresponding to the current feature, and the current feature state is judged based on the confidence corresponding to the current feature and the multi-level dynamic threshold.

[0009] If the current feature status is normal, add a normal feature label and release it.

[0010] If the current feature status is in doubt, a doubtful feature label is added and the review process is triggered.

[0011] If the current feature status is abnormal, it will be marked as a fraud feature label and the review process will be skipped, triggering a fraud alarm.

[0012] A dynamic deception feature database is constructed based on normal feature labels, questionable feature labels and deception feature labels, and the current features are input into the dynamic deception feature database for multimodal deception feature perception.

[0013] The multimodal deception perception results are traced according to the data source ID, and sovereign data isolation is performed based on the tracing results.

[0014] Preferably, constructing a confidence model corresponding to each modal feature specifically includes: The modal features include text modality, image modality and time sequence modality.

[0015] For the text modality, after extracting the semantic vector, a supervised classification model based on semantic consistency is constructed, and the supervised classification model is used as the confidence model corresponding to the text modality; for the image modality, after extracting the feature map, an unsupervised autoencoder model based on reconstruction error is constructed, and the unsupervised autoencoder model is used as the confidence model corresponding to the image modality; for the time series modality, after extracting the behavior coding sequence, a time series classification model based on confidence distribution learning is constructed, and the time series classification model is used as the confidence model corresponding to the time series modality.

[0016] Preferably, a multi-level dynamic threshold is set for each modal feature, specifically including: A normal confidence interval is constructed based on the confidence distribution of historical normal features, and the normal confidence interval boundary is extracted using the quantile statistics method. The normal dynamic threshold is set according to the normal confidence interval boundary to determine the credible range of the current feature.

[0017] By analyzing the confidence fluctuation range of historical deception characteristics, a deception fluctuation interval is constructed, and the interval within the deception fluctuation interval and outside the normal confidence interval is set as a suspicious dynamic threshold.

[0018] The extreme confidence deviation intervals in historical features are extracted through a clustering algorithm, and threshold regression optimization is performed based on the deception feature label verification results to form the limit boundary for distinguishing deception features. The abnormal dynamic threshold is set based on the limit boundary for distinguishing deception features.

[0019] Preferably, the review process specifically includes: The review process performs multimodal cross-validation on the current feature through the confidence model of the remaining modal features. The multimodal cross-validation includes: Based on the modal mapping rule, the current feature is associated with the remaining modal input channels, and the confidence model of the remaining modal features is used to output the confidence value, that is, the cross-validation results of each modal feature are obtained; the cross-validation results of each modal feature are compared with their respective normal dynamic thresholds. If the confidence values ​​of two modal features exceed the normal dynamic threshold range, the current feature is jointly judged as a deceptive feature, added as a deceptive feature label, and the traceability and isolation operations are continued; if less than two modal features exceed the normal threshold range, the questionable feature label of the current feature is cancelled, and it is changed to a normal feature label and released.

[0020] Preferably, a dynamic deception feature database is constructed based on normal feature labels, questionable feature labels, and deception feature labels, specifically including: A trusted feature index area is constructed, where the trusted feature index area is used to record multimodal data corresponding to normal feature tags.

[0021] A modality cross-validation tracking area is constructed, where the modality cross-validation tracking area is used to record multimodal data corresponding to questionable feature labels.

[0022] A high-risk feature archiving area is constructed, where the high-risk feature archiving area is used to record multimodal data corresponding to the deception feature labels.

[0023] A dynamic deception feature database is constructed based on the trusted feature index area, the modal cross-validation tracking area and the high-risk feature archiving area. The label is used as the driving basis for feature classification of the dynamic deception feature database, and the multimodal data corresponding to the normal feature label, the questionable feature label and the deception feature label are respectively divided into corresponding data partitions; the archiving behavior of the normal feature, the questionable feature and the deception feature is automatically triggered by the label status, and the warehousing process completes the feature classification, index mapping and hash identifier generation.

[0024] Preferably, the current feature is input into a dynamic deception feature database for multimodal deception feature perception, specifically including: Calculate the confidence similarity between the current feature and the normal feature in the trusted feature index area.

[0025] Match the confidence fluctuation trajectory of the current feature in the modality cross-validation tracking area.

[0026] The deviation vector of the current feature is retrieved from the high-risk feature archive area.

[0027] The confidence similarity, confidence fluctuation trajectory, and deviation vector are used to determine whether the prediction conditions are met. If the prediction conditions are met, the deception perception result of the current feature is output as a deception feature, triggering subsequent tracing and data isolation operations; otherwise, it is marked as unidentified deception, and the doubtful observation list is updated based on the degree of deviation of the current feature.

[0028] Preferably, the multimodal deception perception results are traced according to the data source ID, and sovereign data isolation is performed based on the traceability results, specifically including: By calling the hash identifier of the corresponding deception feature in the dynamic deception feature database, the data source ID, data collection channel, access device number and initial generation timestamp to which the deception feature belongs are traced to form a complete data source association path; after confirming the validity of the data source ID, other undetermined features uploaded by the data source in the current cycle are automatically retrieved, and a linkage risk investigation is triggered; if batch abnormal features are found in the same source data during the linkage investigation, the sovereign data isolation operation is performed on the data source channel according to the traceability correlation degree.

[0029] The sovereign data isolation operation is triggered by the data source ID in the tracing result. First, it is identified whether the data source in the tracing result is within the sovereign data tag range. If it is within the sovereign tag range, the isolation policy is called to set the upload interface of the data source in the data flow link to read-only or disconnected state; the data access control list is updated synchronously and marked as a risk isolation state, and all subsequent data entry from the source in the risk isolation state is suspended; at the same time, the isolation execution log is reported, including the data source ID, isolation trigger time and corresponding deception feature details.

[0030] The sovereign data leakage prevention method combined with multimodal deception feature perception also includes a suspicious feature re-determination feedback mechanism: If the current feature does not meet the abnormality judgment standard in the multimodal cross-validation and is re-judged as a normal feature, the current feature will continue to be tracked and recorded as a questionable re-observation sample; in the subsequent monitoring cycle, if the current feature is again classified as a questionable state, or has a similar confidence deviation distribution in the modal structure with the confirmed abnormal feature, and meets the re-judgment threshold conditions, the review process will be automatically re-triggered and the extended verification operation will be forced to execute.

[0031] Extended verification includes: dynamically adjusting the number of auxiliary modalities called, enabling the historical deception feature comparison model, performing pattern matching on the current feature and the archived deception samples. If the similarity exceeds the historical deception threshold, the current feature is updated to an abnormal label, and the alarm and traceability operations are performed simultaneously.

[0032] If the current feature does not trigger any abnormal state again within the set observation period, and the similarity with historical deception samples continues to be in a low range, the current feature will be officially transferred to the normal feature whitelist and recorded in the misjudgment-free list of the deception feature database.

[0033] The beneficial effects of this application are: 1. By constructing a multi-level dynamic threshold mechanism based on historical sample distribution, combined with a supervised classification model, an autoencoder model, and a time series discrimination model, a refined distinction between normal, questionable, and abnormal states is achieved. A multimodal cross-validation mechanism and a feedback re-judgment process are introduced to form a closed-loop discrimination system from perception, verification, to review. This effectively reduces the misjudgment rate caused by boundary samples and system noise, overcoming the problems of traditional static judgment methods, such as poor adaptability to distribution changes and lack of a feedback processing mechanism for deceptive features.

[0034] 2. By building a label-driven data archiving structure and a dynamic deception feature database, combined with deception perception, data traceability, linkage investigation, and sovereign data isolation mechanisms, we can achieve full-chain linkage protection from anomaly identification to data source control. This can identify high-risk data sources, trigger targeted isolation, ensure the integrity and controllability of sovereign data in complex attack scenarios, solve the problems of response delay and coarse isolation granularity, and have the ability to continuously learn and expand new deception features.

[0035] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 Flowchart of a sovereign data leakage prevention method combined with multimodal deception feature perception provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0039] See also Figure 1 , Figure 1 A flowchart of a sovereign data leakage prevention method combined with multimodal deception feature perception is provided for an embodiment of the present application.

[0040] In this embodiment, the sovereign data leakage prevention method combined with multimodal deception feature perception includes step S10, step S20, step S30, step S40 and step S50.

[0041] Step S10: extracting features of each modality based on the acquired multimodal data, specifically including: The multimodal data includes text data, image data and time-series behavior data.

[0042] Text data comes from user input records, system logs, communication content or unstructured text.

[0043] Image data comes from monitoring screenshots, interface images, chart data or user-uploaded images.

[0044] Time series behavior data comes from user operation behavior sequences, system access trajectories, input and output records, and information data containing time dimensions.

[0045] Data from different modalities are aggregated on the same data access platform through API calls, and synchronized, standardized, and formatted through a data preprocessing process to ensure alignment and correlation between modalities.

[0046] For text modality, BERT in the pre-trained language model is used to encode the text data, mapping the original sentences or paragraphs into high-dimensional semantic vectors and retaining the contextual semantic features.

[0047] It should be noted that BERT is the abbreviation of Bidirectional Encoder Representations from Transformer. The usage process includes: segmenting the input text, building the input tensor, passing it into the BERT model for encoding, and extracting the average pooling result as the semantic representation.

[0048] For image modality, ResNet in the convolutional neural network is used to extract features from the image and output multi-layer convolution feature maps to capture the spatial structure and local pattern information in the image.

[0049] It should be noted that ResNet stands for Residual Network. Its usage process includes: normalizing the image size and pixels, inputting it into the ResNet network, extracting multi-layer feature outputs through residual connection layers, and ultimately obtaining feature maps. The ResNet network uses a ResNet-50 variant as its base architecture, retaining the residual connection mechanism in the base architecture and replacing the final fully connected layer with an embedded feature output module to generate fixed-dimensional feature representations. In the image preprocessing stage, all images are resized to a 224×224 resolution, normalized to [0, 1], and employ a random cropping and flipping augmentation strategy. During training, the encoder and decoder of the autoencoder model are both symmetrical structures. The encoder compresses to a 128-dimensional embedding vector, uses mean squared error as the reconstruction loss, and uses RMSProp as the optimizer, with an initial learning rate of 0.0005.

[0050] For temporal behavior modes, the LSTM modeling method is used to encode the operation behavior sequence to obtain a dynamic feature representation that can reflect the relationship between behavior order and temporal sequence.

[0051] It should be noted that LSTM is the abbreviation of Long Short-Term Memory. The usage process includes: encoding and embedding the behavior sequence in time steps, inputting it into the LSTM unit, extracting long-term and short-term dependency information through the memory gating mechanism, and outputting a feature vector representing the current sequence state.

[0052] The modal features extracted above will serve as the input basis for subsequent confidence calculation and deception identification; through the above method, unified access and semantic representation of heterogeneous modal data can be achieved, providing multi-dimensional and highly robust feature input guarantees for subsequent feature discrimination, review and verification, and anomaly identification operations.

[0053] Step S20: Construct a confidence model corresponding to each modal feature and set a multi-level dynamic threshold for each modal feature; the multi-level dynamic threshold includes a normal dynamic threshold, a questionable dynamic threshold, and an abnormal dynamic threshold, specifically including: The modal features include text modality, image modality and time sequence modality.

[0054] For text modality, after extracting the semantic vector, a supervised classification model based on semantic consistency is constructed, and the supervised classification model is used as the confidence model corresponding to the text modality.

[0055] It should be noted that the supervised classification model uses a structure that combines a support vector machine and a fully connected neural network, in which the support vector machine is used to construct a primary classifier with optimized boundaries, and the fully connected neural network is used to extract deep semantic features and enhance nonlinear discrimination capabilities. The model training uses a labeled historical text feature set, adopts cross entropy as the loss function, and uses the Adam optimizer in the training process. The initial learning rate value is set to 0.001, the batch size is 64, and the number of training rounds is 20. The positive and negative sample ratio in the training set is maintained at 1:1, and a word vector layer freezing strategy is introduced during the training process to prevent overfitting. The supervised classification model is used to establish a mapping relationship between features and labels by learning the feature boundaries between historically annotated normal text samples and abnormal samples. After the supervised classification model training is completed, the input semantic vector is discriminated and the probability value of the normal feature is output. The obtained probability value is used as the confidence result of the text modality. The supervised classification model can effectively identify whether the text deviates from the existing pattern at the semantic level.

[0056] For the image modality, after extracting the feature map, an unsupervised autoencoder model based on the reconstruction error is constructed, and the unsupervised autoencoder model is used as the confidence model corresponding to the image modality.

[0057] It should be noted that the unsupervised autoencoder model is used to learn the reconstruction ability of image features under unlabeled conditions. The image feature map is input into the encoder for compression, and then the decoder reconstructs the image features. The confidence is evaluated by calculating the reconstruction error between the original feature map and the reconstructed feature map. The smaller the error, the closer the feature is to the historical normal image, and the higher the confidence, otherwise it is low confidence.

[0058] For temporal modalities, after extracting the behavior coding sequence, a temporal classification model based on confidence distribution learning is constructed, and the temporal classification model is used as the confidence model corresponding to the temporal modality.

[0059] It should be noted that the time series classification model is modeled using the LSTM method. The LSTM contains two stacked LSTM units, each unit contains 128 hidden neurons, the maximum length of the input sequence is 200 steps, and the embedding layer dimension is 64. To prevent the gradient from disappearing, a batch normalization layer is introduced and position encoding information is added to the time dimension. The model training uses labeled historical behavior sequences, the loss function is KL divergence, and an early stop strategy is introduced during training to automatically terminate overfitting rounds. The time series classification model is used to learn user behavior sequences. By modeling the state transition pattern and feature distribution of historical normal time series behavior, the matching degree of the input sequence under the learned distribution is calculated, and the corresponding confidence score is output. The time series classification model accurately identifies small offsets or potential abnormal trends in user behavior through time window modeling and distribution deviation analysis.

[0060] It should be noted that multiple levels of dynamic thresholds are set for each modal feature, specifically including: A normal confidence interval is constructed based on the confidence distribution of historical normal features, and the normal confidence interval boundary is extracted using the quantile statistics method. The normal dynamic threshold is set according to the normal confidence interval boundary to determine the credible range of the current feature.

[0061] It should be noted that in the process of constructing the multi-level dynamic threshold of each modality, the probability density estimation of the confidence vector of the historical normal samples is first performed, the kernel density estimation method is used to fit the distribution form of the historical normal samples, and the 95% confidence interval is extracted as the normal judgment boundary; the upper boundary is used to determine the credible release range, and the lower boundary is used as the trigger point of the doubtful state; the multi-level dynamic threshold update adopts a sliding window strategy, and every 1,000 new feature samples are used as the update window. The built-in confidence in the window is resampled and triggers a threshold adaptive adjustment.

[0062] By analyzing the confidence fluctuation range of historical deception characteristics, a deception fluctuation interval is constructed, and the interval within the deception fluctuation interval and outside the normal confidence interval is set as a suspicious dynamic threshold.

[0063] The extreme confidence deviation intervals in historical features are extracted through a clustering algorithm, and threshold regression optimization is performed based on the deception feature label verification results to form the limit boundary for distinguishing deception features. The abnormal dynamic threshold is set based on the limit boundary for distinguishing deception features.

[0064] Step S30: Input the current feature into the corresponding confidence model to calculate the confidence corresponding to the current feature, and determine the current feature state based on the confidence corresponding to the current feature and the multi-level dynamic threshold, specifically including: If the confidence value corresponding to the current feature is within the normal dynamic threshold, it means that the current feature is within the statistical distribution range of historical normal features, and the overall behavior pattern is highly consistent with the existing normal samples, meeting the credible range judgment conditions; at this time, the current feature is determined to be normal, the current feature is added as a normal feature label, and is directly released without entering the subsequent verification process.

[0065] If the confidence value corresponding to the current feature is lower than the lower limit of the normal dynamic threshold but higher than the upper limit of the abnormal dynamic threshold, it indicates that the current feature fails to meet the credible feature judgment standard and does not meet the deviation abnormal judgment condition. There is a certain degree of uncertainty or structural drift. At this time, the current feature is judged to be in a questionable state, and the current feature is added as a questionable feature label. The review process is immediately triggered to assist in judging the true attributes of the current feature through multimodal cross-validation.

[0066] It should be noted that the review process performs multimodal cross-validation on the current feature through the confidence model of the remaining modal features. The multimodal cross-validation includes: Based on the modal mapping rule, the current feature is associated with the remaining modal input channels, and the confidence model of the remaining modal features is used to output the confidence value, that is, the cross-validation results of each modal feature are obtained; the cross-validation results of each modal feature are compared with their respective normal dynamic thresholds. If the confidence values ​​of two modal features exceed the normal dynamic threshold range, the current feature is jointly judged as a deceptive feature, added as a deceptive feature label, and the traceability and isolation operations are continued; if less than two modal features exceed the normal threshold range, the questionable feature label of the current feature is cancelled, and it is changed to a normal feature label and released.

[0067] It should be noted that the modal mapping rules are driven by the multimodal synergy matrix, which is obtained through statistical learning of historical modal synergy judgment samples and can automatically determine the optimal auxiliary modal combination when the current modality is missing. The synergy matrix is ​​updated every 30 data cycles and dynamically adjusts the modal participation order and priority based on the synergy information entropy scoring mechanism. Confidence values ​​are standardized during the cross-validation process to eliminate the impact of inconsistent scales between modalities on the judgment results.

[0068] If the confidence value corresponding to the current feature is less than or equal to the confidence lower limit of the abnormal dynamic threshold, it means that the current feature has seriously deviated from the normal feature distribution and has abnormal performance, and the overall behavioral characteristics are highly similar to the archived deception samples; at this time, the current feature is judged to be in an abnormal state and directly marked as a deception feature label, skipping the review process, immediately triggering the deception alarm operation, and promoting the subsequent data traceability and sovereign data isolation execution process.

[0069] It should be noted that the fraud alert operation is used to complete risk notification and response linkage as soon as the current feature is determined to be a fraud feature, specifically including: Once a deceptive feature label is marked, a real-time alarm is sent immediately. The real-time alarm includes the modal type, confidence value, deviation feature vector, judgment timestamp and data session identifier of the current feature; a structured alarm record is generated synchronously and written to the alarm log for asynchronous retrieval by subsequent audits; during the alarm stage, the data source operation authority is not directly interfered with, but it serves as a pre-risk identification trigger for traceability and isolation logic, providing feature judgment basis and data support for the subsequent sovereign data isolation strategy.

[0070] Through the alarm mechanism, we can achieve early perception, rapid marking and linkage notification of suspicious features, establish a closed-loop risk control path from identification to response, and ensure that sufficient contextual judgment criteria and response basis are in place before isolation execution.

[0071] Step S40: constructing a dynamic deception feature database based on the normal feature labels, the questionable feature labels, and the deception feature labels, and inputting the current feature into the dynamic deception feature database for multimodal deception feature perception, specifically including: A trusted feature index area is constructed, where the trusted feature index area is used to record multimodal data corresponding to normal feature tags.

[0072] A modality cross-validation tracking area is constructed, where the modality cross-validation tracking area is used to record multimodal data corresponding to questionable feature labels.

[0073] A high-risk feature archiving area is constructed, where the high-risk feature archiving area is used to record multimodal data corresponding to the deception feature labels.

[0074] A dynamic deception feature database is constructed based on the trusted feature index area, the modal cross-validation tracking area and the high-risk feature archiving area. The label is used as the driving basis for feature classification of the dynamic deception feature database, and the multimodal data corresponding to the normal feature label, the questionable feature label and the deception feature label are respectively divided into corresponding data partitions; the archiving behavior of the normal feature, the questionable feature and the deception feature is automatically triggered by the label status, and the warehousing process completes the feature classification, index mapping and hash identifier generation.

[0075] It should be noted that the current features are input into the dynamic deception feature database for multimodal deception feature perception, specifically including: Calculate the confidence similarity between the current feature and the normal feature in the trusted feature index area.

[0076] Match the confidence fluctuation trajectory of the current feature in the modality cross-validation tracking area.

[0077] The deviation vector of the current feature is retrieved from the high-risk feature archive area.

[0078] The confidence similarity, confidence fluctuation trajectory, and deviation vector are used to determine whether the prediction conditions are met. If the prediction conditions are met, the deception perception result of the current feature is output as a deception feature, triggering subsequent tracing and data isolation operations; otherwise, it is marked as unidentified deception, and the doubtful observation list is updated based on the degree of deviation of the current feature.

[0079] It should be noted that the prejudgment conditions are: The similarity index of the current feature exceeds the similarity threshold of the normal feature, and the confidence fluctuation trajectory has a fluctuation pattern consistent with the historical questionable features.

[0080] Or, the similarity index of the current feature exceeds the similarity threshold of the normal feature, and the deviation vector exceeds the abnormal deviation judgment threshold.

[0081] Alternatively, the confidence fluctuation trajectory has a fluctuation pattern consistent with historical questionable characteristics, and the deviation vector exceeds the abnormal deviation judgment threshold.

[0082] It should be noted that the similarity threshold of normal features is used to measure the confidence similarity between the current feature and the historical normal features in the trusted feature index area; the mean vector is calculated through the confidence vector set of normal features, and the cosine similarity is used as the measurement method, and the decision boundary is determined by the lower quartile of the historical data; the similarity threshold of normal features is set according to the decision boundary. If the similarity between the current feature and the normal feature is lower than the trusted feature similarity threshold, it is regarded as a feature abnormality and deviates from the normal feature distribution.

[0083] It should be noted that the confidence fluctuation trajectory is obtained by constructing a multi-period confidence sequence and extracting the local fluctuation feature vector in combination with a sliding time window mechanism. The sliding time window mechanism refers to the use of a window with a fixed time span to slide on the sequence in time series analysis, extracting features step by step according to the step size, capturing the local fluctuation pattern and trend changes in the sequence, and is suitable for fluctuation trajectory analysis tasks.

[0084] It should be noted that the identification of the existence of a fluctuation pattern in the confidence fluctuation trajectory that is consistent with the historical questionable features is achieved by recording the confidence values ​​corresponding to the current feature in multiple consecutive judgments in the past, forming a time series sample, and using a fixed-width sliding window to perform mean, variance and fluctuation amplitude analysis, extracting the changing trend of the confidence fluctuation, and performing similarity matching with the historical questionable trajectories in the modal cross-validation tracking area. Based on the similarity matching results, the judgment of whether the trajectory is consistent is completed.

[0085] It should be noted that the abnormal deviation judgment threshold constructs a deviation distribution model through the deviation vector set of historical deception features in the high-risk feature archive area, and then uses the boundary extraction algorithm based on the triple standard deviation principle based on the deviation distribution model to calculate the statistical extreme value boundary of the deviation vector, and sets the abnormal deviation judgment threshold according to the statistical extreme value boundary. The abnormal deviation judgment threshold is used as the judgment standard for whether the current feature deviates from the normal distribution.

[0086] It should be noted that the deviation distribution model refers to aggregating the deviation vectors of historical features, statistically analyzing the distribution characteristics of the deviation vectors of historical features in high-dimensional space, and using clustering methods to identify common offset areas and boundary areas of historical features, and measuring whether new features are out-of-distribution anomalies based on the offset areas and boundary areas.

[0087] Step S50: Tracing the multimodal deception perception results according to the data source ID and performing sovereign data isolation based on the tracing results, specifically including: By calling the hash identifier of the corresponding deception feature in the dynamic deception feature database, the data source ID, data collection channel, access device number and initial generation timestamp to which the deception feature belongs are traced to form a complete data source association path; after confirming the validity of the data source ID, other undetermined features uploaded by the data source in the current cycle are automatically retrieved, and a linkage risk investigation is triggered; if batch abnormal features are found in the same source data during the linkage investigation, the sovereign data isolation operation is performed on the data source channel according to the traceability correlation degree.

[0088] It should be noted that other undetermined features refer to feature samples that are uploaded by the same data source ID as the deception features in the current data cycle, but have not yet completed the complete confidence assessment process or have not yet reached the judgment threshold. Specifically, they include: feature data from which modal features have been extracted but have not yet been labeled; cached feature data that is waiting for evaluation due to resource scheduling and whose judgment results have not yet been returned; and questionable candidate features that have not yet met the judgment conditions in the review process and are still under observation.

[0089] It should be noted that the linkage risk investigation refers to the process of conducting a correlation analysis around all data upload behaviors of the data source in the current cycle to check whether there are data deception characteristics based on confirming the validity of the data source ID of the deception characteristic. Specifically, it includes: Retrieve the homologous feature list, retrieve all undetermined features under the same data source ID, and organize them into an analysis queue in chronological order.

[0090] Trigger the confidence rapid assessment mechanism, and perform confidence assessment and abnormal deviation detection on all undetermined features under the same data source ID in a fast compression mode based on the original confidence model.

[0091] The evaluation results are aggregated and judged, and the abnormal confidence output ratio, deviation amplitude and modal distribution of all undetermined features under the same data source ID are counted to determine whether all undetermined features under the same data source ID meet the abnormal feature conditions.

[0092] Perform linked tagging and feedback storage, mark features that meet abnormal conditions as high-risk candidates, and synchronously update them to the dynamic deception feature database.

[0093] The linkage result drives the isolation trigger. If it is confirmed that there is an abnormal distribution phenomenon of multiple points with the same source, sovereign data isolation will be executed.

[0094] The sovereign data isolation operation is triggered by the data source ID in the tracing result. First, it is identified whether the data source in the tracing result is within the sovereign data tag range. If it is within the sovereign tag range, the isolation policy is called to set the upload interface of the data source in the data flow link to read-only or disconnected state; the data access control list is updated synchronously and marked as a risk isolation state, and all subsequent data entry from the source in the risk isolation state is suspended; at the same time, the isolation execution log is reported, including the data source ID, isolation trigger time and corresponding deception feature details.

[0095] In addition, it should be noted that the "sovereign data" mentioned in this specification and claims refers to data resources that are legally acquired and owned by enterprises, institutions or organizations and can be collected, transmitted, used and analyzed within the scope of internal control; typical forms of sovereign data include business operation logs, platform interaction data, internal behavior records, process flow data and anonymous user usage tracks. The circulation scope is controlled by the data management entity and is controllable, auditable and legal.

[0096] The sovereign data involved in this invention does not include sensitive data or personal privacy data in the legal sense. The leakage prevention strategy implemented is aimed at the internal compliance management needs of the enterprise and the goal of maintaining data integrity. It aims to ensure the safe and controllable use of sovereign data within the authorized boundaries through technical means, and prevent the risk of abnormal data leakage caused by technical deception.

[0097] The sovereign data leakage prevention method combined with multimodal deception feature perception also includes a suspicious feature re-determination feedback mechanism: If the current feature does not meet the abnormality judgment standard in the multimodal cross-validation and is re-judged as a normal feature, the current feature will continue to be tracked and recorded as a questionable re-observation sample; in the subsequent monitoring cycle, if the current feature is again classified as a questionable state, or has a similar confidence deviation distribution in the modal structure with the confirmed abnormal feature, and meets the re-judgment threshold conditions, the review process will be automatically re-triggered and the extended verification operation will be forced to execute.

[0098] Extended verification includes: dynamically adjusting the number of auxiliary modalities called, enabling the historical deception feature comparison model, performing pattern matching on the current feature and the archived deception samples. If the similarity exceeds the historical deception threshold, the current feature is updated to an abnormal label, and the alarm and traceability operations are performed simultaneously.

[0099] If the current feature does not trigger any abnormal state again within the set observation period, and the similarity with historical deception samples continues to be in a low range, the current feature will be officially transferred to the normal feature whitelist and recorded in the misjudgment-free list of the deception feature database.

[0100] It should be noted that the set observation period refers to the time window for continuous monitoring of features marked as "doubtful re-observation samples". The length of the set observation period is dynamically adjusted based on the actual operating frequency and the modal data update rate. The default period is nearly three data refresh cycles. During the set observation period, the features marked as "doubtful re-observation samples" will be continuously perceived and modal fluctuations analyzed.

[0101] It should be noted that the re-judgment threshold condition means that the current feature shows suspicious modal deviation again during the continuous observation process, and the confidence similarity, deviation distance or fluctuation amplitude between the current feature and the historical abnormal feature simultaneously meet the preset joint indicator judgment conditions, including: the confidence value of the current feature is lower than the credible boundary twice in a row; the average cosine similarity between the fluctuation trajectory of the current feature and the deviation trajectory of the historical deception feature is higher than 0.85; the proportion of the current feature appearing in the two types of high-risk modalities exceeds 50%.

[0102] It should be noted that the dynamic adjustment of the number of auxiliary modalities refers to the automatic increase or decrease of the number of modalities involved in the analysis according to the historical deviation records of the current features during the extended verification phase. The text modality, image modality and time series modality are called by default.

[0103] It should be noted that the historical deception feature comparison model is a multimodal pattern matching model constructed based on archived deception features. It integrates convolutional neural networks and an attention mechanism to construct an inter-modal alignment embedding space, which is used to compare the comprehensive similarity of the current feature in terms of structure, confidence, and deviation trend, and outputs a similarity score result. The historical deception feature comparison model uses a Transformer structure to construct a cross-modal unified feature space, in which the image modality is mapped to the same dimension as the text modality after ResNet encoding, and a modality fusion representation is constructed through a cross-modal attention mechanism. During the training phase, the model uses triplet loss optimization to maximize the similarity of the current feature with similar deception samples and minimize the similarity with normal samples. The training set consists of archived deception samples and whitelisted samples without false positives. The training epoch is set to 100, the optimizer is AdamW, the initial learning rate is 1e-4, and the learning rate is dynamically decayed using a cosine annealing scheduler.

[0104] It should be noted that the historical deception threshold refers to the minimum similarity judgment limit between the current feature and the archived deception sample. If the comprehensive similarity score between the current feature and the historical deception sample exceeds the historical deception threshold, the current feature is considered to be consistent with the historical deception pattern, which constitutes the basis for re-marking.

[0105] It should be noted that the false positive avoidance list is a specific partition in the dynamic deception feature database, which is used to register feature samples that have been repeatedly marked as normal during multiple evaluations, have experienced short-term suspicious fluctuations but were eventually confirmed to be normal. The false positive avoidance list automatically reduces the probability of suspicious features being mislabeled as abnormal in future similar feature evaluations.

[0106] In addition, to verify the feasibility of the technical solution of this application, an experimental comparison was conducted between the technical solution of this application and another type of privacy protection method for multimodal data in the disclosed technology. This experiment used a simulated dataset covering three modalities: text, image, and temporal behavior. Each modality included normal and abnormal samples to comprehensively evaluate the performance of the two technical solutions. The specific experimental data is as follows: Text Data Modality: Generates normal text snippets as normal data, while simulating text containing sensitive keywords, encoded information, or disguised formats as abnormal samples to represent potential data leakage text patterns. Text features are represented by extracting semantic vectors to evaluate the technical solution's ability to detect abnormal semantic content.

[0107] Image data modality: Regular images are collected as normal samples, and images with embedded hidden watermarks are simulated as abnormal samples. Image features are represented by extracting feature matrices to test the recognition of image leakage behavior.

[0108] Time series behavior modality: Generates normal user behavior sequences as normal data, and simulates abnormal behavior sequences as abnormal data. Time series data extracts behavioral feature sequences for analysis, simulating possible data leakage behavior patterns. Time series behavior modality is used to evaluate the response speed to deceptive behavior and the detection robustness.

[0109] The above simulated data covers multimodal feature dimensions, ensuring that each modality has a certain proportion of abnormal samples; at the same time, a cross-modal combination scenario is constructed: some abnormal samples only behave abnormally in a single modality, while other abnormal samples have slight deviations in multiple modalities. In order to test the multimodal cross-validation capability of the new method and the performance differences of the existing method when distinguishing each modality separately, abnormal samples account for about 10% of the overall data set, ensuring that the false alarm rate and missed alarms in the evaluation process are statistically significant.

[0110] To ensure a fair, clear, and reasonable comparison, this experiment conducted comparable tests on the two technical solutions under the same environment. The process is as follows: A1. Model initialization: Use the same historical normal data to train / configure the sub-methods. For the technical solution provided in this application, confidence models for the three modalities of text, image, and time series are constructed respectively, and their respective dynamic threshold ranges are set according to the normal sample distribution; for the privacy protection method for another category of multimodal data in the public technology, a public detection mechanism is used for configuration: For the sake of fairness, the privacy protection method for another category of multimodal data in the public technology is also based on the static threshold of each modality obtained through training as the abnormality judgment standard, but no cross-modal verification is performed. This ensures that the two methods face the same training conditions, and differ only in the threshold discrimination mechanism and multimodal collaboration.

[0111] A2. Test Data Injection: The same simulated test data sequence is fed into both models. Data samples from the test set are fed one by one in the order they were generated, including a mixed sequence of normal and abnormal samples. To simulate a real-time detection environment, the timestamp of each data sample input is recorded for subsequent calculation of recognition latency.

[0112] A3. Anomaly detection process. After receiving the input, the two methods each execute anomaly determination logic: This application extracts features from the input multimodal data, calculates the confidence level and first makes a judgment based on the dynamic threshold of each modality. If a modal feature is lower than the normal threshold, it is marked as "doubtful", triggering a multimodal cross-verification process: that is, the confidence model of the remaining modalities is called to verify the same data. If at least two modal results exceed the normal range threshold, the comprehensive judgment is made as abnormal, otherwise the doubtful judgment is cancelled and released; if a modality directly exceeds the abnormal threshold, it is immediately judged as abnormal and an alarm is triggered without review; this process aims to improve detection accuracy and reduce false alarms by utilizing dynamic threshold adaptability and cross-modal consistency.

[0113] Another type of privacy protection method for multimodal data in the public technology compares the characteristics of each modality of the input data with a static threshold. Because this technical solution lacks a cross-modal verification mechanism, if any modality exceeds the static threshold, the input is judged to be abnormal and an alarm is issued; if all modalities are within the threshold, it is considered to have passed normally. A single judgment strategy may mistakenly identify random fluctuations in certain modalities as abnormalities. The lack of precise distinction between "abnormal characteristics" and "deceptive characteristics" leads to a high false alarm rate. It may also lack sensitivity to minor anomalies in multimodal collaboration. Static rules are difficult to adapt to the distribution differences under different modal combinations.

[0114] A4. Index recording: For each input, the detection results and timestamp are recorded in real time.

[0115] The record content includes: the true attributes of the sample (normal or abnormal), the judgment result given by the detection method (normal pass or abnormal alarm), and the time when the alarm is triggered. Subsequent indicators are calculated based on the records: If the sample is a true anomaly and the technical solution determines it to be an anomaly, it is counted as a successful detection and the recognition delay (alarm time minus sample input time) is calculated.

[0116] If the sample is a true anomaly but the method does not raise an alarm, it is counted as a missed alarm.

[0117] If the sample is truly normal but the method is incorrect and the alarm is triggered, it will be recorded as a false alarm; the time point when the false alarm occurs is recorded for reference in delay statistics (normal samples should not trigger alarms. This experiment also records the delay when the false alarm is triggered but it is not counted in the "abnormal identification delay" indicator).

[0118] Correct release of normal samples and correct interception of abnormal samples are both counted as correct identification times and used to calculate the overall accuracy or specific accuracy indicators.

[0119] A5. To enhance the reliability of the experimental conclusions, the above test process was repeated 10 times independently using the same parameter settings. Each iteration randomly regenerated the simulated sample details to ensure that the results were independent of a specific dataset. After each round of experiments, the values ​​of the three indicators were recorded, and the results of the 10 rounds were averaged to obtain the final average performance indicator. Average statistics avoid smoothing out accidental fluctuations, making comparisons more objective. For each indicator, if the standard deviation across multiple experiments was less than 5% of the original value, the results were statistically significant, confirming that the performance difference was due to the technical solution itself rather than random factors.

[0120] This experiment compares the effects of the two methods based on the three key performance indicators mentioned above. The definition, calculation method and significance of each indicator are explained as follows: Recognition accuracy: This is used to measure the ability to correctly identify abnormal data. Recognition accuracy = the number of successfully identified abnormal samples / the total number of abnormal samples × 100%. A higher ratio indicates that the method is able to capture more real anomalies, that is, the false negative rate is low, reflecting the effectiveness of the model in detecting attack behaviors.

[0121] Anomaly Identification Latency: This measures the time between when data input is identified as an anomaly and an alert is issued. The identification latency is calculated for each successfully detected anomaly sample, and the average of all anomaly sample detection latencies is used as the metric. The calculation formula is: Anomaly Identification Latency = Σ(Anomaly Alert Trigger Time - Anomaly Sample Input Time) / Number of Detected Anomaly Samples. A shorter latency indicates a faster response, enabling earlier detection and prevention of data leaks.

[0122] False Alarm Rate: This measures the proportion of normal data mistakenly identified as abnormal, reflecting detection accuracy. It is calculated as: False Alarm Rate = (Number of Falsely Alarmed Normal Samples / Total Number of Normal Samples) × 100%. A lower false alarm rate indicates higher accuracy, meaning that false alarms for normal behavior are less frequent, thus minimizing disruption to normal business operations.

[0123] After averaging 50 tests using the same test data and environment, the average performance indicators of the two solutions are shown in the following table:

[0124] It should be noted that another category of privacy protection methods for multimodal data in the public technology specifically refers to privacy protection methods based on verification of transformed data generation and verification mechanisms. The privacy protection methods based on verification of transformed data generation and verification mechanisms are to generate transformed data in text / image form by extracting key features, and setting up two verification procedures to confirm user access rights; the privacy protection methods based on verification of transformed data generation and verification mechanisms are characterized by the use of only static thresholds and lack of adaptability to dynamic changes in abnormal evolution; lack of collaborative judgment capabilities between cross-modal features, resulting in a high false alarm rate and a high risk of missed detection; and inability to cope with complex scenarios of slight modal drift and weak feature attacks.

[0125] Through comparative experiments based on simulated samples and combined with the data in the above table, it can be seen that: although another category of multimodal data solutions in the public technology can detect some abnormal data in a single mode, because the core mechanism relies on static thresholds and fixed rule combinations, it is difficult to adapt to the dynamic changes of abnormal features and cross-drift between feature modes in the actual environment; especially when there is slight interference or cross-modal camouflage features in the multimodal data, the other category of multimodal data solutions in the public technology are prone to increased false alarm rates and insensitivity to deceptive feature recognition in anomaly detection; in contrast, the technical solution provided by this application has achieved effective improvements and enhancements in multiple key capabilities: A multimodal confidence model is introduced to establish independent confidence distributions for text, image and temporal behavior modes respectively, and multi-level dynamic thresholds are set based on historical normal features, with the ability to automatically adjust the judgment interval as the data evolves; a "doubtful state" intermediate judgment mechanism is set to avoid direct misjudgment of boundary samples, and a multimodal cross-verification strategy is used to jointly re-judge the sample state based on the consistent response of other modalities, thereby improving the overall recognition robustness; a dedicated deception feature archiving database is constructed, and historical deception feature similarity matching is called in the review process to achieve deep re-identification and intelligent traceability of abnormal features, thereby enhancing the ability to handle fuzzy attacks or policy evasion behaviors.

[0126] Comprehensive comparative experiments have confirmed that the technical solution provided by this application has excellent performance in the three core indicators of recognition accuracy, response delay and false alarm rate, especially in complex sample scenarios and multimodal mild drift, with stronger adaptability and defense capabilities; the relevant performance improvement comes from the targeted optimization of the technical solution provided by this application in terms of algorithm structure, data adaptation and judgment strategy, which effectively compensates for the static rigidity, rough judgment and frequent false alarms of existing methods, and provides a deployable, traceable and sustainably optimized technical path for the prevention and control of sovereign data leakage.

[0127] At this point, the sovereign data leakage prevention method combined with multimodal deception feature perception is completed.

[0128] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process is determined by the function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0129] Those skilled in the art will appreciate that the various algorithms or steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0130] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A sovereign data leakage prevention method based on multimodal deception feature perception, characterized in that: These include: Extracting features of each modality based on the acquired multimodal data; Constructing a confidence model corresponding to each modal feature and setting a multi-level dynamic threshold for each modal feature; the multi-level dynamic threshold includes a normal dynamic threshold, a questionable dynamic threshold, and an abnormal dynamic threshold; Input the current feature into the corresponding confidence model to calculate the confidence corresponding to the current feature, and judge the current feature state based on the confidence corresponding to the current feature and the multi-level dynamic threshold; If the current feature status is normal, add a normal feature label and release it; If the current feature status is in doubt, add a doubtful feature label and trigger the review process; If the current feature status is abnormal, it will be marked as a fraud feature label and the review process will be skipped, triggering a fraud alarm; A dynamic deception feature database is constructed based on normal feature labels, questionable feature labels, and deception feature labels, and the current feature is input into the dynamic deception feature database for multimodal deception feature perception. The multimodal deception perception results are traced according to the data source ID, and sovereign data isolation is performed based on the tracing results.

2. The sovereign data leakage prevention method combined with multimodal deception feature perception as claimed in claim 1 is characterized in that: Construct the confidence model corresponding to each modal feature, including: The modal features include text modality, image modality and time sequence modality; For the text modality, after extracting the semantic vector, a supervised classification model based on semantic consistency is constructed, and the supervised classification model is used as the confidence model corresponding to the text modality; for the image modality, after extracting the feature map, an unsupervised autoencoder model based on reconstruction error is constructed, and the unsupervised autoencoder model is used as the confidence model corresponding to the image modality; for the time series modality, after extracting the behavior coding sequence, a time series classification model based on confidence distribution learning is constructed, and the time series classification model is used as the confidence model corresponding to the time series modality.

3. The sovereign data leakage prevention method combined with multimodal deception feature perception as claimed in claim 2 is characterized in that: Set multi-level dynamic thresholds for each modal feature, including: Based on the confidence distribution of historical normal features, a normal confidence interval is constructed. The normal confidence interval boundary is extracted using the quantile statistical method. The normal dynamic threshold is set according to the normal confidence interval boundary to determine the credible range of the current feature. By analyzing the confidence fluctuation range of historical deception characteristics, a deception fluctuation interval is constructed, and the interval within the deception fluctuation interval but outside the normal confidence interval is set as a suspicious dynamic threshold; The extreme confidence deviation intervals in historical features are extracted through a clustering algorithm, and threshold regression optimization is performed based on the deception feature label verification results to form the limit boundary for distinguishing deception features. The abnormal dynamic threshold is set based on the limit boundary for distinguishing deception features.

4. The sovereign data leakage prevention method combined with multimodal deception feature perception as claimed in claim 1 is characterized in that: The review process specifically includes: The review process performs multimodal cross-validation on the current feature through the confidence model of the remaining modal features. The multimodal cross-validation includes: Based on the modal mapping rule, the current feature is associated with the remaining modal input channels, and the confidence model of the remaining modal features is used to output the confidence value, that is, the cross-validation results of each modal feature are obtained; the cross-validation results of each modal feature are compared with their respective normal dynamic thresholds. If the confidence values ​​of two modal features exceed the normal dynamic threshold range, the current feature is jointly judged as a deceptive feature, added as a deceptive feature label, and the traceability and isolation operations are continued; if less than two modal features exceed the normal threshold range, the questionable feature label of the current feature is cancelled, and it is changed to a normal feature label and released.

5. The sovereign data leakage prevention method combined with multimodal deception feature perception as claimed in claim 1 is characterized in that: A dynamic deception feature database is constructed based on normal feature labels, suspicious feature labels, and deception feature labels, specifically including: Constructing a trusted feature index area, wherein the trusted feature index area is used to record multimodal data corresponding to normal feature tags; Constructing a modality cross-validation tracking area, wherein the modality cross-validation tracking area is used to record multimodal data corresponding to the questionable feature label; Constructing a high-risk feature archiving area, wherein the high-risk feature archiving area is used to record multimodal data corresponding to the deception feature labels; A dynamic deception feature database is constructed based on the trusted feature index area, the modal cross-validation tracking area and the high-risk feature archiving area. The label is used as the driving basis for feature classification of the dynamic deception feature database, and the multimodal data corresponding to the normal feature label, the questionable feature label and the deception feature label are respectively divided into corresponding data partitions; the archiving behavior of the normal feature, the questionable feature and the deception feature is automatically triggered by the label status, and the warehousing process completes the feature classification, index mapping and hash identifier generation.

6. The sovereign data leakage prevention method combined with multimodal deception feature perception as claimed in claim 5 is characterized in that: Input the current features into the dynamic deception feature database for multimodal deception feature perception, specifically including: Calculate the confidence similarity between the current feature and the normal feature in the trusted feature index area; Match the confidence fluctuation trajectory of the current feature in the modality cross-validation tracking area; Retrieving the deviation vector of the current feature in the high-risk feature archive area; The confidence similarity, confidence fluctuation trajectory, and deviation vector are used to determine whether the prediction conditions are met. If the prediction conditions are met, the deception perception result of the current feature is output as a deception feature, triggering subsequent tracing and data isolation operations; otherwise, it is marked as unidentified deception, and the doubtful observation list is updated based on the degree of deviation of the current feature.

7. The sovereign data leakage prevention method combined with multimodal deception feature perception as claimed in claim 1 is characterized in that: Multimodal deception perception results are traced according to the data source ID, and sovereign data isolation is performed based on the traceability results, including: By calling the hash identifier of the corresponding deception feature in the dynamic deception feature database, the data source ID, data collection channel, access device number and initial generation timestamp to which the deception feature belongs are traced, forming a complete data source association path; after confirming the validity of the data source ID, other undetermined features uploaded by the data source in the current cycle are automatically retrieved, and a linkage risk investigation is triggered; if batch abnormal features are found in the same source data during the linkage investigation, the sovereign data isolation operation is performed on the data source channel based on the traceability association degree; The sovereign data isolation operation is triggered by the data source ID in the tracing result. First, it is identified whether the data source in the tracing result is within the sovereign data tag range. If it is within the sovereign tag range, the isolation policy is called to set the upload interface of the data source in the data flow link to read-only or disconnected state; the data access control list is updated synchronously and marked as a risk isolation state, and all subsequent data entry from the source in the risk isolation state is suspended; at the same time, the isolation execution log is reported, including the data source ID, isolation trigger time and corresponding deception feature details.

8. The sovereign data leakage prevention method combined with multimodal deception feature perception as claimed in claim 1, characterized in that: It also includes a feedback mechanism for re-determining questionable features: If the current feature fails to meet the abnormality judgment criteria in the multimodal cross-validation and is re-judged as a normal feature, the current feature will be tracked and recorded as a suspicious re-observation sample; in the subsequent monitoring cycle, if the current feature is again classified as a suspicious state, or has a similar confidence deviation distribution in the modal structure with the confirmed abnormal feature, and meets the re-judgment threshold conditions, the review process will be automatically re-triggered and the extended verification operation will be forced to execute; Extended verification includes: dynamically adjusting the number of auxiliary modalities called, enabling a historical deception feature comparison model, matching the current feature with archived deception samples, and updating the current feature to an anomaly label if the similarity exceeds the historical deception threshold. Alerts and tracebacks are also performed simultaneously. If the current feature does not trigger any abnormal state again within the set observation period, and the similarity with historical deception samples continues to be in a low range, the current feature will be officially transferred to the normal feature whitelist and recorded in the misjudgment-free list of the deception feature database.

Images