Multi-modal data dynamic desensitization method and system based on machine learning

Abstract

The invention relates to the technical field of information security management, in particular to a multi-modal data dynamic desensitization method and system based on machine learning, multi-modal features such as texts, images and time sequences are jointly extracted through CNN-Transform-LSTM, 1-5 sensitive levels are evaluated by adopting a cross-modal attention model, a self-adaptive desensitization strategy is generated through an improved MOEA / D-DE algorithm, and the dynamic desensitization of the multi-modal data is realized. And realizing cross-modal collaborative desensitization by combining GNN. The system dynamically adjusts parameters through TD3 reinforcement learning, the response delay is less than or equal to 100ms, the Bi-LSTM-AE model evaluates privacy and utility in real time, and the block chain evidence storage whole process is realized. The method solves the problems of precision and utility imbalance, poor dynamic adaptability and insufficient multi-mode collaboration in the traditional technology, and is suitable for the fields of finance, medical treatment and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data security and privacy protection, and particularly relates to a multi-modal data dynamic desensitization method and system based on machine learning. BACKGROUND

[0002] In the digital era, data as a core production factor, its privacy protection and safe use of the contradiction is increasingly prominent. The wide application of multi-modal data (text, image, audio, etc.) exacerbates the risk of privacy leakage, but the existing desensitization technology has three core bottlenecks:

[0003] (1) Desensitization precision and data utility imbalance

[0004] Traditional rule-based desensitization (such as fixed character replacement, sensitive area deletion) adopts a "one-size-fits-all" strategy, which cannot adapt to the semantic features of data. For example: in the financial field, a bank desensitizes the customer flow text, simply replacing the "account number" "amount" and other keywords, resulting in the destruction of semantic coherence, and the accuracy of subsequent risk control models for identifying abnormal transactions decreases from 90% to 52%; in the medical field, the sensitive area (such as the patient's face) of the CT image is pixel-level blurred, although the privacy is protected, the spatial position correlation of the lesion and the face is lost, resulting in a 40% increase in the misdiagnosis rate of the auxiliary diagnosis system. The core defect of such methods is that they do not distinguish between "sensitive features" and "utility features" in the data, resulting in a significant decrease in the analysis value of the desensitized data, with a utility loss of more than 40%.

[0005] (2) Insufficient adaptability to dynamic scenarios

[0006] The sensitivity level of data changes dynamically with the use scenario, access subject, and time, but traditional static desensitization strategies cannot respond in real time: medical data needs to retain high-precision lesion information (sensitivity level S = 2) in the "clinical diagnosis" scenario, but needs to weaken individual identification (sensitivity level S = 4) in the "scientific research statistics" scenario, but the traditional method switching strategy requires manual configuration, with a response delay of more than 24 hours; logistics data can show complete recipient information (S = 1) in the "internal scheduling" scenario, but needs to hide address details (S = 3) when sharing externally, static desensitization leads to "overprotection" or "insufficient protection", increasing the risk of privacy leakage by 35%. The existing technology lacks the ability to perceive scene characteristics and dynamically adjust strategies in real time, and cannot meet the needs of dynamic privacy protection.

[0007] (3) Difficulty in collaborative desensitization of multi-modal data

[0008] Heterogeneous data such as text, images, audio, etc. exist cross-modal correlation (such as "express single text address" and "recipient portrait"), but the traditional method adopts independent desensitization process, leading to associated privacy leakage: the e-commerce platform desensitizes the commodity comment text (hides the mobile phone number) and the buyer's portrait (blurs the processing) respectively, but the attacker can locate the user through the cross of "delivery address" in the text and "community background" in the portrait, and the associated leakage risk increases by 28%; in the government system, the identity card image desensitization still retains the "date of birth" digital feature, and the linkage with the "age" information in the text form can reverse the complete identity card number, and such vulnerabilities are due to the lack of modeling and collaborative protection of cross-modal correlation.

[0009] In summary, due to the three problems of "precision-utility imbalance", "poor dynamic adaptability" and "insufficient collaboration", the prior art is difficult to meet the privacy protection needs of multi-modal data in multiple scenarios, and it is urgent to build an intelligent desensitization system integrating machine learning. SUMMARY

[0010] The present application aims to provide a multi-modal data dynamic desensitization method based on machine learning, which is suitable for the whole life cycle privacy protection of structured data (such as transaction flow, medical record form) and unstructured data (such as text comments, medical images, voice recordings) in multiple scenarios such as finance, medical treatment, logistics, etc. It can realize accurate desensitization and utility balance of cross-modal data, meet the compliance requirements of GDPR, etc. Protection level four, etc., specifically including the following steps:

[0011] S1 Multi-modal data feature joint extraction, extract image visual features by CNN, extract text semantic features by Transformer, extract audio time sequence features by LSTM, and extract time sequence data patterns by TCN, and construct a cross-modal feature matrix; S2 Dynamic evaluation of sensitive level, based on three-dimensional features of data content, use scene and access subject, output 1-5 level sensitive level through cross-modal attention model;

[0012] S3 Self-adaptive desensitization strategy generation, using improved MOEA / D-DE algorithm to build a multi-objective optimization framework to generate a Pareto optimal strategy set;

[0013] S4 Cross-modal collaborative desensitization control, model the modal correlation relationship through graph neural network, realize joint optimization of desensitization parameters; S5 Dynamic desensitization execution and feedback, based on TD3 reinforcement learning algorithm to adjust desensitization parameters in real time, combined with Bi-LSTM-AE model to evaluate desensitization effect.

[0014] Further, the weight calculation formula of the cross-modal attention model is:

[0015]

[0016] wherein v is the image feature vector, t is the text feature vector, s is the time series feature vector, Wq, Wk are learnable weight matrices.

[0017] Further, the improved MOEA / D-DE algorithm comprises a dynamic neighborhood adjustment mechanism, and the formula is:

[0018] wherein D(t) is the current neighborhood size, f(t) is the population convergence degree, and λ=0.05 is the decay coefficient.

[0019] Further, the node state update formula of the graph neural network is:

[0020]

[0021] wherein is the l-th layer feature of node v, N(v) is the adjacent node set, and cvu is the associated edge weight.

[0022] Further, the fusion time consumption of the multi-modal data feature joint extraction is ≤50 ms, wherein the image feature adopts an improved ResNet-50 model to output a 2048-dimensional vector, and the text feature adopts a BERT-base model to output a 768-dimensional vector.

[0023] Further, the reward function of the TD3 reinforcement learning algorithm is:

[0024] R = 0.6 · (S old -S new ) - 0.3 · U loss - 0.1 · C

[0025] wherein Sold-Snew is the sensitivity level before and after desensitization, Uloss is the utility loss rate, and C is the calculation overhead.

[0026] Further, the Bi-LSTM-AE evaluation model simultaneously outputs the privacy entropy and the utility loss rate, and triggers the strategy reconstruction when the utility loss rate exceeds 20%.

[0027] Further, it further comprises a blockchain full-link storage step, adopts a consortium chain architecture to record the desensitization parameters, timestamps, and operation subjects, and automatically verifies the compliance through a smart contract.

[0028] Further, the system of the method comprises:

[0029] A multi-modal data access layer supports text, image, audio, and time series data collection;

[0030] An intelligent feature processing layer integrates a cross-modal attention evaluation module and a graph neural network collaborative framework;

[0031] A dynamic strategy generation layer includes an MOEA / D-DE optimizer and a TD3 reinforcement learning unit;

[0032] A desensitization execution layer deploys a differential privacy, a homomorphic encryption and a neuron masking processor;

[0033] A central control platform stores desensitization records and supports full-link visualized tracing.

[0034] Further, the system, the intelligent feature processing layer includes a sensitive level evaluation module, and the central control platform integrates a blockchain storage module.

[0035] Beneficial effects:

[0036] The cross-modal attention evaluation model greatly improves the sensitive level judgment accuracy, can accurately identify the associated sensitive features such as “text name-image face”, significantly reduces the omission rate, and solves the mixed modal data protection blind area problem. The improved MOEA / D-DE algorithm effectively improves the generation efficiency of the optimal desensitization strategy, realizes the balance of “high desensitization strength + high utility reservation” in the financial scene, and improves the semantic integrity of the input data of the risk control model.

[0037] The GNN cross-modal collaborative framework reduces the risk of text-image cross-modal association leakage through graph node association modeling, reduces the user privacy complaints of e-commerce platforms, and at the same time guarantees the accuracy of cross-modal data analysis. The TD3 reinforcement learning model greatly shortens the sensitive level response delay, significantly improves the strategy adjustment efficiency in the medical data scene switching, and the dynamic protection capability can meet the real-time business needs.

[0038] The Bi-LSTM-AE evaluation model reduces the desensitization effect judgment error, improves the utility loss over-standard early warning accuracy, and avoids the data value loss caused by “over desensitization”. The neuron masking algorithm preserves key semantic features in image desensitization, improves the identity card OCR recognition rate, and reduces the misdiagnosis rate of medical image auxiliary diagnosis. The blockchain full-link tracing mechanism shortens the audit time, reduces the annual compliance cost, at the same time meets the requirements of GDPR and etc. four-level protection, and improves the cross-border data cooperation scene expansion capability. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The overall flowchart of the present application;

[0040] Figure 2 Data processing flow; Figure 3 Data desensitization flow; Figure 4 Data optimization flow. DETAILED DESCRIPTION

[0041] Embodiment 1:

[0042] A machine learning-based multi-modal data dynamic de-sensitization method, comprising the following steps:

[0043] 1. Multi-modal data feature joint extraction

[0044] Heterogeneous data access: Collect text (such as transaction logs, medical records), images (such as ID cards, CT images), audio (such as customer service recordings), and time series data (such as sensor signals) through edge computing nodes, and perform format standardization (text segmentation, image resizing, audio sampling rate unified to 16kHz).

[0045] Feature extraction network:

[0046] Image features: Use an improved ResNet-50 model to strengthen sensitive area (such as face, lesion) features through attention mechanism, output 2048-dimensional visual feature vector;

[0047] Text features: Use BERT-base model to extract context semantic features, construct domain word table for financial terms (such as "credit" "overdue") and medical terms (such as "tumor" "heart rate"), output 768-dimensional text feature vector;

[0048] Audio features: Use a combination of LSTM+CNN models, first extract acoustic features through Mel-spectrum conversion, then capture temporal dependencies through LSTM, output 512-dimensional audio feature vector;

[0049] Time series data features: Use TCN (Temporal Convolution Network) to extract periodic patterns, output 128-dimensional time series feature vector.

[0050] Cross-modal fusion: Through feature concatenation and attention weighting (weights dynamically adjusted through training), construct a unified cross-modal feature matrix, feature discrimination improved by 50% compared to single-modal extraction, fusion time ≤ 50ms.

[0051] 2. Dynamic evaluation of sensitive level

[0052] Three-dimensional feature input:

[0053] Data content features: Based on the cross-modal feature matrix of step 1, extract the confidence of sensitive entities (such as ID number, lesion location);

[0054] Usage scenario features: Convert scenario labels (such as "internal audit" "public sharing") and preset rules into scenario vectors;

[0055] Access subject features: Based on subject permission levels (such as "administrator" "third party") and historical operation records, generate trustworthiness vectors.

[0056] Evaluation Model: A Transformer-based cross-modal attention evaluation model is constructed, which dynamically weights three-dimensional features through self-attention mechanisms and outputs 1-5 sensitive levels (S=1 is the lowest and S=5 is the highest). The evaluation accuracy reaches 98.3%.

[0057] Real-time update: The input features are refreshed every 500ms to realize dynamic tracking of sensitive levels, with a response speed 100 times faster than traditional rule methods.

[0058] 3. Adaptive desensitization strategy generation

[0059] Multi-objective optimization framework: Three optimization objectives are defined:

[0060] Desensitization strength (α): The mapping of sensitive level S and desensitization algorithm strength (e.g., S=5 corresponds to differential privacy ε=0.1);

[0061] Data utility (β): The performance retention rate of data after desensitization in target tasks (such as classification, retrieval);

[0062] Computational overhead (γ): Time and resource consumption of desensitization operation;

[0063] Constraint condition: α+β+γ=1, dynamically adjusted by weight (e.g., β weight increased to 0.5 in financial scenarios).

[0064] Strategy optimization algorithm: Improved MOEA / D-DE algorithm:

[0065] Dynamic neighborhood adjustment mechanism is introduced, and the neighborhood size changes adaptively with the convergence degree of the population, with the formula:

[0066] where D(t) is the current neighborhood size, f(t) is the population convergence degree, λ

[0067] is the decay coefficient (value 0.05);

[0068] Generate a Pareto optimal strategy set (including differential privacy, homomorphic encryption, neuron masking, etc. Algorithm combination), and the strategy generation efficiency is improved by 60% compared with NSGA-III algorithm.

[0069] 4. Cross-modal collaborative desensitization control

[0070] Graph neural network modeling:

[0071] Node definition: Abstract text fields (such as "address"), image regions (such as "face"), and audio segments (such as "name pronunciation") as graph nodes;

[0072] Edge weight calculation: generate edge weight based on association rule mining (e.g. "text address and image landmark co-occurrence frequency") to represent the strength of modality association (range 0-1).

[0073] Co-optimization: cross-modality parameter joint optimization through GNN message passing mechanism, node state update formula:

[0074] where is the l-th layer feature of node v, N(v) is the adjacent node set, and cvu is the associated edge weight, ensuring the desensitization strength of the associated nodes to be collaborative (e.g. when "address text" is desensitized, "landmark image" is simultaneously enhanced for desensitization), reducing the risk of privacy leakage by 75%.

[0075] 5. Dynamic desensitization execution and feedback

[0076] Desensitization engine: integrate multiple desensitization algorithms:

[0077] Text: BERT-based neuron masking (masking the neurons corresponding to sensitive words);

[0078] Image: adaptive pixel blur (sensitive area blur radius dynamically adjusted with S);

[0079] Audio: LSTM-based speech feature replacement (preserve intonation, replace acoustic features of sensitive words).

[0080] Reinforcement learning tuning: use TD3 (Double Delayed Deep Deterministic Policy Gradient) algorithm:

[0081] State space: sensitivity level S, current desensitization parameters, data utility indicators;

[0082] Action space: algorithm selection and parameter adjustment (e.g. differential privacy ε ± 0.1);

[0083] Reward function:

[0084] R = 0.6 · (S old -S new ) - 0.3 · U loss - 0.1 · C, where Sold-Snew is the difference in sensitivity level before and after desensitization, Uloss is the utility loss, and C is the computation overhead.

[0085] Through the double critic network and delay update mechanism, avoid Q value overestimation, parameter adjustment delay ≤100ms, stability improved by 80% compared with DDPG algorithm.

[0086] 6. Intelligent evaluation of desensitization effect

[0087] Evaluation model: build a combined model of Bi-LSTM and autoencoder (AE):

[0088] Bi-LSTM captures the temporal / semantic coherence of the data after desensitization (such as text sentence structure, audio tone change);

[0089] AE reconstructs the original data and evaluates the feature retention degree through reconstruction error;

[0090] Output privacy entropy (measure the strength of privacy protection) and utility loss rate, evaluate the accuracy rate of 95%.

[0091] Feedback mechanism: when the utility loss exceeds the threshold (such as 20%), trigger policy reconstruction, return to step 3 to generate an optimized policy.

[0092] 7. Full-link desensitization trajectory tracing

[0093] Blockchain storage: adopt consortium chain architecture (nodes include data parties, user parties, and regulatory parties), record:

[0094] Desensitization operation log (parameters, timestamps, operation subjects);

[0095] Sensitive level evaluation results and policy generation process;

[0096] Effectiveness evaluation report and abnormality early warning record.

[0097] Smart contract: predefine compliance rules (such as "medical data desensitization needs to retain lesion features"), automatically verify the compliance of desensitization operations, when detecting violations (such as excessive desensitization leading to excessive utility loss), block the operation in real time and trigger audit, trace delay reduced to within 1 second, efficiency improved by 10 times compared with traditional log systems.

[0098] The present application also provides an intelligent desensitization system for implementing the above method, comprising:

[0099] Multi-modal data access layer: supports text (TXT, JSON), image (JPG, DICOM), audio (WAV, MP3) and other formats, integrates edge nodes for preprocessing (denoising, format conversion), throughput ≥100MB / s.

[0100] Intelligent feature processing layer: deploy CNN-Transformer-LSTM joint extraction module (feature extraction accuracy 98%), cross-modal attention evaluation engine (sensitive level evaluation accuracy 98.3%), and GNN collaborative framework (correlation modeling error ≤5%).

[0101] Dynamic policy generation layer: includes MOEA / D-DE optimizer (policy generation time ≤200ms), TD3 reinforcement learning unit (parameter adjustment delay ≤100ms), and policy knowledge base (stores ≥100,000 historical policies).

[0102] Desensitization execution layer: integrated differential privacy module (epsilon range 0.1-10), homomorphic encryption engine (supports Paillier algorithm), neuron mask processor (adapts to mainstream deep learning framework), execution delay ≤ 50ms.

[0103] Central control platform: with data visualization (real-time display of desensitization success rate, utility loss rate), policy management (manual intervention interface), effect evaluation and blockchain storage functions, storage ≥ 100 million desensitization records, supporting 7x24 hour operation

[0104] Firstly, the Transformer attention mechanism is introduced into multi-modal sensitive evaluation. Through dynamically weighting the three-dimensional features of data content, scene and subject, the accurate division of 1-5 level sensitive grades is realized, and the evaluation accuracy reaches 98.3%, which is 46% higher than the traditional rule method (accuracy 52%). Especially in mixed modal data (such as "text + image"), through self-attention to capture cross-modal association (such as the correspondence between "text name" and "image face"), the problem of missing judgment of associated sensitive features in traditional methods is solved.

[0105] Improved MOEA / D-DE multi-objective desensitization strategy algorithm

[0106] A dynamic neighborhood adjustment mechanism is proposed, which improves the convergence speed of the Pareto optimal strategy set by 55% compared with the NSGA-III algorithm, and the optimal strategy search efficiency is improved by 70%. By balancing the desensitization strength, data utility and computational overhead, the optimal solution of "high desensitization strength (α=0.6) + high utility retention (β=0.35)" is realized in the financial scene, solving the problem of "insufficient protection" or "excessive utility loss" caused by traditional single-objective optimization.

[0107] GNN cross-modal collaborative desensitization framework

[0108] The association between multi-modal data (such as "express text address" and "avatar background") is modeled using graph neural networks, and the joint optimization of cross-modal desensitization parameters is realized through the node state update formula, reducing the risk of privacy leakage by 75%. Compared with the independent desensitization process, this framework can automatically identify high-risk association paths (edge weight ≥ 0.7) and trigger collaborative protection, filling the technical gap in cross-modal privacy protection.

[0109] TD3 dynamic desensitization reinforcement learning model

[0110] A double-critic network and a delayed update mechanism are designed. In the scenario of real-time change of sensitivity level, the desensitization parameter adjustment delay is shortened from 500 ms of the traditional DDPG algorithm to 100 ms, and the policy stability is improved by 80%. In the scenario of switching of medical data, the policy adaptation from "clinical" to "scientific" is realized within seconds, and the problem of response lag of the traditional method is solved.

[0111] Bi-LSTM-AE desensitization effect evaluation model

[0112] The Bi-LSTM-AE desensitization effect evaluation model combines bidirectional long short-term memory network and autoencoder, and can capture both the time sequence features (such as speech tone) and the structural features (such as text semantics) of the data. The evaluation error is reduced by 38% compared to the single AE model. It can output privacy entropy and utility loss rate in real time, providing accurate feedback for policy optimization and avoiding "blind desensitization".

[0113] Neuron masking dynamic desensitization algorithm

[0114] Based on deep learning neuron importance analysis, an interpretable neuron-level desensitization method (such as masking the neuron output corresponding to the "mobile phone number" in the BERT model) is developed. In image desensitization, 85% of the key semantic features (such as lesion edges) are preserved, and the data utility is improved by 60% compared to traditional pixel-level desensitization, solving the problem of loss of analysis value caused by "over-desensitization".

[0115] Blockchain-smart contract full-link traceability mechanism

[0116] The desensitization whole process (evaluation, policy generation, execution, evaluation) is stored on the chain for evidence, and the smart contract automatically verifies the desensitization compliance. The traceability efficiency is improved by 10 times compared to the traditional log system, meeting the requirements of GDPR and the fourth level of network security protection. In financial auditing, the desensitization operation is traceable and cannot be tampered with, and the auditing time is shortened from 3 days to 2 hours.

[0117] Embodiment 2

[0118] Financial multi-modal data intelligent desensitization application

[0119] A bank processes 100,000 transactions per day, including:

[0120] Text data: customer transaction records (including account number, transaction amount, and counterparty information);

[0121] Image data: ID card front and back, bank card photo;

[0122] Audio data: customer service call recording (including customer name and password prompt).

[0123] There are three major problems in traditional desensitization solutions: OCR recognition rate of identity card image desensitization decreases from 95% to 45%, affecting automatic review; text desensitization leads to a 60% decrease in semantic analysis accuracy; cross-modal association leakage (such as "water address" and "avatar background") triggers privacy complaints.

[0124] Technical implementation steps

[0125] Multi-modal feature joint extraction

[0126] Text processing: BERT-base model is used to extract the semantic features of the water text, focusing on capturing the context association of entities such as "account number" and "amount" (such as the dependency relationship between "5 million" and "transfer" in "transfer amount 5 million");

[0127] Image processing: Improved ResNet-50 model is used to extract identity card image features, and attention mechanism is used to strengthen the feature expression of "name" and "ID number" area (heat map shows that the area weight is 0.7);

[0128] Audio processing: LSTM+CNN combined model is used to extract acoustic features of the call recording, and sensitive speech fragments such as "password" and "bank card number" are identified;

[0129] Feature fusion: cross-modal feature matrix is constructed through attention weighting (text 0.4, image 0.4, audio 0.2), and the fusion time is 45ms.

[0130] Dynamic assessment of sensitive level

[0131] Three-dimensional feature input:

[0132] Data content: identity card number recognition confidence 0.95 (high), amount entity confidence 0.8 (medium);

[0133] Scene: "internal audit" scene (label converted to vector [1, 0]);

[0134] Subject: bank teller (permission level 3 / 5);

[0135] Model output: sensitive level S = 2 (medium), updated every 500ms (such as switching to "external sharing" scene, S automatically increases to 4).

[0136] Adaptive strategy generation

[0137] Multi-objective optimization: set weight α = 0.3 (desensitization strength), β = 0.5 (utility), γ = 0.2 (overhead);

[0138] MOEA / D-DE algorithm generates strategy set:

[0139] Optimal strategy: "neuron masking" for text (hiding account numbers), "local blur" for images (5px blur radius for ID number area), and "feature replacement" for audio (replacing sensitive speech segments with acoustic features);

[0140] Strategy generation time: 180ms, 64% faster than traditional methods (500ms).

[0141] Cross-modal collaborative desensitization control

[0142] GNN modeling: associate "streaming text address" (node T1) with "ID address" (node I1), and the edge weight is calculated as 0.8 (strong association);

[0143] Collaborative strategy: T1 desensitization (hide street door number), I1 synchronous enhanced desensitization (blur address area), avoid cross positioning by attackers.

[0144] Dynamic desensitization execution and evaluation

[0145] TD3 parameter tuning: when the sensitivity level increases from S=2 to S=4 (external sharing scenario), the policy network outputs "image blur radius increases to 10px" and "text mask range expands", with a delay of 90ms;

[0146] Effect evaluation: Bi-LSTM-AE model calculates utility loss 12% (lower than threshold 20%), privacy entropy 0.8 (high protection strength), no need to reconstruct the strategy.

[0147] Blockchain traceability

[0148] Evidence content: desensitization parameters (blur radius 5px), timestamp (2023-10-01 09:30:00), and operation subject (clerk ID:001);

[0149] Smart contract verification: verified by the rule "ID desensitization needs to retain the outline", marked as compliant operation.

[0150] Implementation effect

[0151] Precision and utility balance:

[0152] After ID image desensitization, the OCR recognition rate remains 85% (traditional method 45%), and the automatic audit pass rate increases from 52% to 90%;

[0153] After the desensitization of the streaming text, the semantic analysis accuracy rate is 88% (traditional 52%), and the false alarm rate of abnormal transaction detection decreases by 70%.

[0154] Dynamic scenario adaptation:

[0155] The policy adjustment delay is 120ms when the scenario switches from "internal audit" (S=2) to "external sharing" (S=4), which is 72000 times higher than manual configuration (24 hours);

[0156] The privacy leakage risk is reduced from 35% to 3%, and no leakage event occurs due to scenario switching.

[0157] Cross-modal collaborative protection:

[0158] The text-image association leakage risk is reduced from 28% to 7%, and the number of user privacy complaints is reduced by 82%;

[0159] Cross-modal data analysis (such as "streaming + avatar" user portrait) still maintains an accuracy of 75%, supporting precise marketing and other businesses.

[0160] Compliance and traceability:

[0161] Blockchain storage reduces audit time from 3 days to 2 hours, and annual audit cost is reduced by 4 million yuan;

[0162] Through the fourth level of network security and GDPR certification, cross-border financial data cooperation business is expanded.

[0163] Embodiment 3:

[0164] Cross-border e-commerce multi-modal desensitization cluster

[0165] The cross-border e-commerce platform deploys the system in the present invention in 20 data centers around the world to process data such as product reviews (text), buyer avatars (images), and customer service calls (audio), and needs to meet multiple regional regulations such as GDPR (European Union) and CCPA (California).

[0166] Technical innovation implementation

[0167] Federal learning collaboration: each data center locally trains a desensitization model and only shares model parameters to avoid cross-border transmission of raw data;

[0168] Cross-chain storage: use blockchain cross-chain technology (such as Polkadot) to realize the interconnection of desensitization records of different regional nodes, and meet regional compliance requirements;

[0169] Dynamic regulatory adaptation: preset regional rules through smart contracts (such as the "right to be forgotten" required by GDPR, which automatically triggers complete data deletion), and prioritize local regulations when generating policies.

[0170] Implementation effect

[0171] Cross-border data flow efficiency is improved by 30% while meeting multiple regional compliance requirements;

[0172] Cross-modal correlation leakage risk dropped from 32% to 8%, and international user satisfaction increased by 40%.

Claims

1. A machine learning-based multi-modal data dynamic de-sensitization method, characterized in that, Comprise the following steps: S1 multi-modal data feature joint extraction, through CNN to extract image visual features, transformer to extract text semantic features, LSTM to extract audio timing features and TCN to extract timing data patterns, to construct a cross-modal feature matrix; S2 dynamic evaluation of sensitivity level, based on three-dimensional features of data content, use scene and access subject, through a cross-modal attention model to output a sensitivity level of 1-5; S3 adaptive desensitization strategy generation, an improved MOEA / D-DE algorithm is used to construct a multi-objective optimization framework to generate a Pareto optimal strategy set, the improved MOEA / D-DE algorithm includes a dynamic neighborhood adjustment mechanism, the formula is: where D(t) is the current neighborhood size, f(t) is the population convergence degree, and λ = 0.05 is the decay coefficient. S4 cross-modal collaborative desensitization control, the modal correlation is modeled by a graph neural network to realize joint optimization of desensitization parameters; S5 dynamic desensitization execution and feedback, based on TD3 reinforcement learning algorithm to adjust desensitization parameters in real time, combined with Bi-LSTM-AE model to evaluate desensitization effect.

2. The method of claim 1, wherein, The weight calculation formula of the cross-modal attention model is: Where v is the image feature vector, t is the text feature vector, s is the timing feature vector, and Wq and Wk are learnable weight matrices.

3. The method of claim 1, wherein, The node state update formula of the graph neural network is: wherein is the l-th layer feature of node v, N(v) is the set of neighboring nodes, and cvu is the associated edge weight.

4. The method of claim 1, wherein, The fusion time of the multi-modal data feature joint extraction is ≤50ms, wherein the image feature uses an improved ResNet-50 model to output a 2048-dimensional vector, and the text feature uses a BERT-base model to output a 768-dimensional vector.

5. The method of claim 1, wherein, The reward function of the TD3 reinforcement learning algorithm is: R = 0.6 · (S old - S new - 0.3 · U loss - 0.1 · C Where Sold-Snew is the sensitivity level before and after desensitization, Uloss is the utility loss rate, and C is the calculation overhead.

6. The method of claim 1, wherein, The Bi-LSTM-AE evaluation model outputs privacy entropy and utility loss rate simultaneously, and triggers strategy reconstruction when the utility loss rate exceeds 20%.

7. The method of claim 1, wherein, It also includes a blockchain full-link storage step, which uses a consortium chain architecture to record desensitization parameters, timestamps and operation subjects, and automatically verifies compliance through a smart contract.

8. A system for implementing the method of any one of claims 1 to 7, characterized in that, Comprise: Multi-modal data access layer, supporting text, image, audio and timing data collection; Intelligent feature processing layer, integrating cross-modal attention evaluation module and graph neural network collaborative framework; Dynamic strategy generation layer, including MOEA / D-DE optimizer and TD3 reinforcement learning unit; Desensitization execution layer, deploying differential privacy, homomorphic encryption and neuron mask processor; Central control platform, storing desensitization records, supporting full-link visual traceability.

9. The system of claim 8, wherein, The intelligent feature processing layer includes a sensitivity level evaluation module, and the central control platform integrates a blockchain storage module.

Images