A privacy enhancement and compliance management method for mental health high-sensitive data

Abstract

The application discloses a privacy enhancement and compliance management method for mental health high-sensitive data, relates to the technical field of artificial intelligence, data security and privacy computing, and comprises the following steps: acquiring original data of a psychological consultation scene; for the data scene of the original data, the original data is processed by using a privacy enhancement mode corresponding to the data scene, so that first training data is obtained; a to-be-trained model is trained by using the first training data, and first weight update parameters of the to-be-trained model are obtained based on a DP-SGD algorithm; the to-be-trained model is adjusted based on the first weight update parameters, so that a trained inference model is obtained; an inference result of the inference model is obtained, and if abnormal information exists in the inference result, the inference result is intercepted, so that the whole-process management and control of model training can be realized.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, data security, and privacy computing, and in particular to a method for enhancing privacy and managing compliance for highly sensitive data related to mental health. Background Technology

[0002] In the field of mental health AI (Artificial Intelligence), data is the most sensitive of all personal information. Existing data processing solutions typically rely on general single-point desensitization tools, which cannot accurately identify complex entities unique to psychological counseling scenarios, such as implicit kinship relationships, medical institutions combined with medical history, etc., and lack the ability to control the entire process from data access, training to prevent inference, to output security alignment.

[0003] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art. Summary of the Invention

[0004] The purpose of this invention is to provide a method for enhancing privacy and managing compliance of highly sensitive mental health data, thereby enabling full-process control of model training.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for enhancing privacy and managing compliance of highly sensitive mental health data includes: Obtain raw data from psychological counseling scenarios; For the data scenario of the original data, the original data is processed using the privacy enhancement method corresponding to the data scenario to obtain the first training data; The first training data is used to train the model to be trained, and the first weight update parameters of the model to be trained are obtained based on the DP-SGD algorithm. The model to be trained is adjusted based on the first weight update parameter to obtain the trained inference model. Input data is provided to the trained inference model to obtain the inference result of the inference model. If abnormal information is detected in the inference result, the inference result is intercepted. The input data is multimodal medical data.

[0006] Optionally, the method further includes: A five-level sensitivity classification model is used to label the sensitivity level and data category of the original data.

[0007] Optionally, the method further includes: Generate an interactive electronic authorization contract that matches the sensitivity level of the original data; The interactive electronic authorization contract is encrypted and stored using a blockchain.

[0008] Optionally, the privacy enhancement methods include: Perform a hash operation on the original data; Alternatively, the original data can be replaced using a compliant generalization algorithm based on K-anonymity; Alternatively, noisy data can be added to the original data based on differential privacy technology.

[0009] Optionally, before training the model using the first training data, the method further includes: A re-identification attack simulation is performed on the first training data. If the simulation result indicates that the de-identification of the first training data has failed, the privacy enhancement processing is performed on the first training data again. And / or, perform semantic availability verification on the first training data. If the verification result indicates that the information entropy loss of the first training data is too high, then perform noise reduction processing on the first training data.

[0010] Optionally, the model to be trained is a local model; adjusting the model to be trained based on the first weight update parameters to obtain the trained inference model includes: The first weight update parameter is sent to the central server so that the central server can calculate the third weight update parameter based on the first weight update parameter and the second weight update parameter obtained from other training devices. The training model is adjusted by receiving the third weight update parameters sent by the central server to obtain the trained inference model.

[0011] Optionally, before the central server updates the parameters according to the first weight and the second weight obtained from other training devices, the method further includes: The first weight update parameter is split based on a secure multi-party computation protocol to obtain a first parameter ciphertext, so that the central server can calculate the third weight update parameter based on the first parameter ciphertext and the second parameter ciphertext obtained from other training devices based on the split second weight update parameter.

[0012] Optionally, the method further includes: A call log for the first training data is generated, and the hash digest of the call log is stored on the blockchain.

[0013] An electronic device includes a processor and a memory, wherein a computer program is stored in the memory, and when executed by the processor, the computer program implements any of the above-described methods for enhancing privacy and managing compliance of highly sensitive mental health data.

[0014] A readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described methods for enhancing privacy and managing compliance of highly sensitive mental health data.

[0015] In the solution provided by this invention, before model training, the original data is processed using an appropriate privacy enhancement method based on the data scenario, thereby providing secure training data. During training, the DP-SGD algorithm is used for gradient pruning to further reduce privacy data leakage during training. After training, when the model is used for inference, abnormal information is intercepted, achieving true closed-loop control throughout the entire process and solving the problem of non-compliant large model training data from the source. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for enhancing privacy and managing compliance of highly sensitive mental health data, provided as an embodiment of the present invention; Figure 2 A flowchart illustrating a federated model training method provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating a privacy data processing method provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of a permission isolation architecture provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the embodiments of the present invention. Please refer to the drawings to make the objectives, features, and advantages of the present invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the implementation conditions of the present invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by the present invention, should still fall within the scope of the technical content disclosed in the present invention.

[0018] In one embodiment of the present invention, see Figure 1 This paper provides a method for enhancing privacy and managing compliance of highly sensitive mental health data, including the following steps S101-S105.

[0019] S101: Obtain raw data from psychological counseling scenarios; S102: For the data scenario of the original data, the privacy enhancement method corresponding to the data scenario is used to process the original data to obtain the first training data; S103: Train the model to be trained using the first training data, and obtain the first weight update parameters of the model to be trained based on the DP-SGD algorithm; S104: Adjust the model to be trained based on the first weight update parameters to obtain the trained inference model; S105: Provide input data to the trained inference model to obtain the inference result of the inference model. If abnormal information is detected in the inference result, the inference result is intercepted. The input data is multimodal medical data.

[0020] In one embodiment, a five-level sensitivity classification model can be used to label the sensitivity level and data category of the original data.

[0021] For raw data, we can obtain its data classification, which can specifically include five classification dimensions: identity identification, health diagnosis and treatment, psychological content, scene association, and generalized statistical information.

[0022] In this case, the sensitivity level can be marked as follows: Level L1 (Very Low Sensitivity / Public Generalized Data): Data containing only the "Generalized Statistics" dimension.

[0023] Level L2 (Low Sensitivity / Scene-Related Data): Contains only data from a single "scene-related" or "occupation-related" dimension, and does not contain any health or identity information.

[0024] Level L3 (Medium-sensitivity / Objective Health Data): Includes independent "health diagnosis and treatment" dimension data, but does not contain specific identity references.

[0025] Level L4 (Highly Sensitive / Deep Psychological Content): Contains highly private "psychological content" dimension data, such as psychometric features.

[0026] Level L5 (Extremely Sensitive / Strongly Correlated Top-Secret Data): Data in which the "identity identifier" dimension co-occurs with any one or more other dimensions, or composite data that can be accurately inferred from a specific natural person through scene combinations.

[0027] Specifically, a Named Entity Recognition (NER) algorithm based on BERT (Bidirectional Encoder Representations from Transformers) can be used, combined with a regular expression rule engine, to identify the original data and obtain data classification.

[0028] In this case, the system's built-in five-level sensitivity classification model can be used to classify the raw data of each data category. After classification, different privacy enhancement methods can be implemented according to the five-level classification strategy.

[0029] Determining the data scenario: such as Figure 2 This system utilizes a dedicated entity recognition model for psychological counseling to accurately cover complex, scenario-based sensitive information, such as precise identification of professions and employers. Specifically, it acquires scenario-based sensitive information as raw data, maps various entities extracted from this information into feature vectors, and inputs these feature vectors into an attention mechanism layer to calculate semantic tensor similarity. Based on the calculated semantic tensor similarity, it determines the combined information entropy between features. If the combined information entropy is lower than a preset safety threshold, it is considered to have an extremely high identification risk, and the sensitivity level of the sensitive information is forcibly increased, thereby achieving anti-inference protection in the early stages of training.

[0030] Specifically, the data scenario is identification information containing one or more dimensions of data.

[0031] For example, if "scene association" dimension data and "health diagnosis and treatment" dimension data are extracted from sensitive information, the former has a sensitivity level of L2 and the latter has a sensitivity level of L3 / or L4. However, if the combined information entropy of the feature vectors of the two is lower than the security threshold, it means that the combination of the "scene association" dimension data and the "health diagnosis and treatment" dimension data can infer the specific individual identity. In this case, the data containing this combination is treated as a whole, and the sensitivity level of this combination is set to L5. Thus, the highest level of isolation and desensitization strategy can be directly applied in subsequent processing.

[0032] To identify various entities, the entity recognition model constructs a local co-occurrence graph network with text entities as nodes. If entities of high-risk dimensions are detected to co-occur in the same context, the model will trigger a "clustering effect of associated nodes" to dynamically upgrade the ranking.

[0033] In one embodiment, the entity recognition model calculates the overall sensitivity score S as follows: in, Indicates the basic privacy weight; Indicates the confidence level of the feature; Indicates the co-occurrence penalty coefficient; Indicates a co-occurrence indicator function; This refers to the unique identifiers of entities identified in the text, including but not limited to patient names, ID card numbers, employee numbers, contact numbers, email addresses, precise physical addresses, and specific job titles, which can directly locate or point to a specific natural person. This refers to the deep psychological and medical content entity features identified in the text, including but not limited to specific scores of psychological tests, scale results, records of chief complaints in psychological counseling, psychiatric diagnosis history, crisis intervention records, medication prescriptions, and highly sensitive medical and psychological professional information such as veiled expressions of traumatic experiences.

[0034] The co-occurrence indicator function is a binary indicator function used to determine whether the aforementioned identity-type entity features and content-type entity features appear simultaneously within a set distance (preset context window). If so, then... =1; in other cases =0.

[0035] The preset context window can be set to the same natural paragraph, the same consultation dialogue sentence group, or two entities separated by no more than a preset number of tokens in the text (such as a span of 50 characters). As long as both entities are detected to appear simultaneously within the window, the function value is set to 1, thereby triggering a step weighting of the (co-occurrence penalty coefficient) in the comprehensive scoring formula.

[0036] Feature confidence comes from the classification probability score output by the last layer of the Softmax or Sigmoid activation function when the entity recognition model performs feature classification prediction. This value, between 0 and 1, represents the confidence level of the entity recognition model in determining that the text fragment belongs to a specific sensitive level entity. The entity recognition model can be a NER (Named Entity Recognition) model based on BERT (Bidirectional Encoder Representations from Transformers).

[0037] Among them, the co-occurrence penalty coefficient and the basic privacy weight are preset coefficients. For the basic privacy weight, the higher the sensitivity level, the higher the value of the basic privacy weight. For example, the basic privacy weight of the "identity identifier" level is greater than that of the "psychological content" level, and the basic privacy weight of the "psychological content" level is greater than that of the "health diagnosis and treatment" level.

[0038] The aforementioned entity recognition model ultimately outputs the final sensitivity level and data classification of the original data, which is then attached to the original data as metadata tags, seamlessly connecting to downstream compliance management.

[0039] The sensitivity level is determined based on the comprehensive sensitivity score S. Specifically, the comprehensive sensitivity score S serves as the mapping basis for automatic grading. The system internally presets discretized scoring threshold ranges corresponding to the five sensitivity levels (L1 to L5). After the entity recognition model calculates the continuous comprehensive sensitivity scores S, it automatically and accurately determines and outputs the final sensitivity level of the data by aligning and mapping the S values ​​with the preset threshold ranges. For example, the threshold limits include [T1, T2, T3, T4], and when S > T4, the sensitivity level is marked as L5.

[0040] Furthermore, the aforementioned comprehensive sensitivity score S can serve as a quantitative lever for dynamic upgrades and penalty triggering. When a co-occurrence penalty term is triggered in the formula... When the calculated score S jumps, the score will undergo a numerical step. Once the score S after the step jump crosses the upper threshold limit, the system will trigger a dynamic upgrade and forced adjustment mechanism based on the change in score S, thus avoiding the rigid omissions caused by simply relying on static dictionary-based rating.

[0041] The overall sensitivity score S can also be used as an adjustment parameter for downstream differentiated privacy enhancement strength selection, that is, as a parameter input to the downstream metadata. For example, in the third case of privacy enhancement in the following embodiments, the privacy budget parameter of injected noise can be dynamically and finely adjusted according to the specific value of S within the same sensitivity level. That is, the higher the score S, the greater the intensity of injected noise, realizing seamless linkage between quantitative scoring and desensitization algorithm strength throughout the entire process.

[0042] Compliance management includes privacy enhancement methods in the following three scenarios, as well as the establishment of a collaborative zero-trust architecture.

[0043] First, operators need to perform zero-trust dynamic identity verification and least privilege verification.

[0044] Under the interconnected zero-trust architecture, hierarchical labels are directly bound to the three partitions of the underlying system's three-tiered permission isolation architecture (original sensitive data operation area, de-identified data model development area, and management-level de-identified decision-making macro-decision area). The zero-trust dynamic identity verification module will perform dynamic identity verification and least privilege verification based on the hierarchical results. By constructing an abnormal behavior real-time detection model, it will isolate or automatically block unauthorized / illegal access behaviors and generate full-node operation logs. Normal behaviors are encrypted and stored in the blockchain's immutable evidence storage network, supporting compliance review and judicial evidence for financing.

[0045] The entity recognition model for psychological counseling can be trained using sample data with labeled scenes. The loss is calculated based on the difference between the predicted scene output by the entity recognition model and the labeled scene, and the model parameters are adjusted based on the calculated loss to complete the training.

[0046] In this case, different privacy enhancement methods can be predefined for different scenarios, professions, and organizations. like Figure 3 As shown, in the first case, for data categorized as "identity identifier," the original data can be hashed. Specifically, a salted hash algorithm such as SHA-256+Salt (SHA: Secure Hash Algorithm) can be used to map variable-length plaintext names to fixed-length ciphertext, mathematically guaranteeing the irreversibility of the computation, completely severing the direct connection between the data and the specific natural person, and achieving irreversible hash desensitization.

[0047] In the second scenario, for data categorized as "scenario / occupational association," the original data can be compliantly generalized and replaced using the K-anonymity algorithm. The replacement employs a "Generalization Hierarchy Tree," where the algorithm automatically replaces high-granularity information by moving up the tree structure. For example, "27 years old" is replaced with "20-30 years old range," and "specific company" is replaced with "a certain technology industry." This preserves non-individual features crucial for AI training while preventing hackers from pinpointing specific individuals.

[0048] In the third scenario, for data categorized as "psychological statistical features," noise can be added to the original data based on differential privacy techniques. Specifically, noise can be added to achieve differential privacy (DP) perturbation. Differential privacy is a high-order mathematical privacy protection technique. For continuous numerical values ​​(such as a specific score of 85 on a psychological test), the system does not output the original value directly, but instead injects random mathematical noise conforming to a Laplace or Gaussian distribution. For example, after adding noise, the score of 85 becomes 83.2 or 87.5, but when thousands of data points are aggregated for model training, the macroscopic statistical characteristics (such as mean and variance) remain unchanged. This prevents hackers from deducing the exact score of a specific patient while ensuring the effectiveness of the training.

[0049] The process of training the model to be trained includes: inputting the first training data, forward propagation of the model to calculate the prediction, calculating the loss, backpropagation to calculate the gradient, and finally the optimizer updating the parameters according to the rules.

[0050] The process of obtaining the inference result of the inference model includes: providing input data to the trained inference model, enabling the inference model to output based on the learned knowledge. For example, the scene category information of psychological counseling content is labeled on the first training data, and the training model outputs the scene category judgment as the inference result.

[0051] The DP-SGD (Differentially Private Stochastic Gradient Descent) algorithm is used to perform gradient clipping during the gradient parameter update process of the model to be trained, so as to obtain the first weight update parameters.

[0052] The specific model to be trained can be a large model such as natural language processing or LLaMA (Large Language Model MetaAI).

[0053] like Figure 2 As shown, the first weight update parameters are encrypted and uploaded, that is, the gradient weight encrypted parameters are uploaded to the central server.

[0054] Gradient pruning can prevent the neural network of the model to be trained from memorizing a patient's specific chat history word for word because the network is too large.

[0055] Furthermore, the DP-SGD algorithm can also protect against member inference attacks and data extraction attacks.

[0056] Methods for intercepting inference results include: performing compliance pre-verification during the model inference output stage, using a security rule engine to intercept and reconstruct generated content that may leak personal information or violate medical compliance, thereby achieving a closed-loop control system for the entire AI chain.

[0057] For example, by using a pre-trained abnormal behavior early warning model for identification, when the system detects that the inference model is about to answer a sentence containing privacy leaks or violations, it will "intercept and discard" the sentence in milliseconds and force the output of replacement information, such as: "Sorry, to protect privacy, I cannot provide this information."

[0058] Abnormal behavior early warning models can be trained in advance using sample corpora labeled with abnormal information to acquire recognition capabilities.

[0059] In one embodiment, an interactive electronic authorization contract matching the sensitivity level of the original data can be generated; the interactive electronic authorization contract is then encrypted and stored using a blockchain.

[0060] Specifically, matching can be performed as follows: based on the sensitivity level, different authorization terms, usage purposes, and limiting conditions are automatically invoked. In this case, an interactive contract construction method can be adopted, whereby the data provider (authorizer) and the data receiver (authorized party) confirm parameters such as purpose and duration through an interactive UI interface, and an executable electronic contract is generated in real time, such as a blockchain smart contract based on Hyperledger Fabric or Ethereum.

[0061] Dynamically match the data usage purpose restrictions corresponding to the sensitivity level, generate interactive electronic authorization contracts, and encrypt and store them on the blockchain to ensure that the source of subsequent large model training data is absolutely legal, and retain key information through blockchain storage.

[0062] In one embodiment, such as Figure 3 As shown, after fusing the desensitized data, before training the model to be trained using the first training data, a built-in dual verification mechanism can be automatically executed, including: performing a re-identification attack simulation on the first training data; if the simulation result indicates that the desensitization of the first training data has failed, then performing privacy enhancement processing on the first training data again. And / or, perform semantic availability verification on the first training data; if the verification result indicates that the information entropy loss of the first training data is too high, then perform noise reduction processing on the first training data. The implementation method for re-identification attack simulation is shown below.

[0063] The system internally simulates a hacker algorithm that attempts to cross-reference and match external public databases. For example, it combines keywords such as "20-30 years old," "a certain technology company," and "severe anxiety" to match external public databases, determining whether a specific individual can be identified. The probability of the guess is used as the simulation result. If the probability of guessing exceeds the safety threshold (e.g., 0.01), the system determines that the desensitization has failed and automatically increases the generalization strength or noise of the previous step.

[0064] Semantic usability verification includes calculating the information loss before and after de-identification. An information loss threshold can be set; if the information loss exceeds this threshold, it indicates excessive loss. In this case, a warning message can be output indicating over-de-identification, and the noise parameters of differential privacy perturbations can be dynamically adjusted to maximize the retention of the psychological semantic information required for AI training. Figure 3 Calculation of psycho-semantic retention in the center.

[0065] The dual verification performed in the above embodiments refers to a two-way balance assessment of the "privacy and security" and "AI training availability" of the data. Insufficient anonymization can lead to leaks, while excessive anonymization can render the data unusable. For example, turning all characters in the original data into asterisks would prevent the model from learning. Therefore, semantic availability verification is performed to obtain a highly usable anonymized dataset as the first training data.

[0066] Therefore, in the solution provided by this embodiment of the invention, before model training, the original data is processed using a suitable privacy enhancement method based on the data scenario, thereby providing secure training data. During training, the DP-SGD algorithm is used for gradient pruning to further reduce privacy data leakage during training. After training, when the model is used for inference, abnormal information is also intercepted, achieving true closed-loop control throughout the entire process and solving the problem of non-compliant large model training data from the source.

[0067] In one embodiment, if there is a need for joint modeling, joint training is performed on local models on multiple devices through a central server, as detailed below.

[0068] If the model to be trained is a local model, the first weight update parameter is sent to the central server so that the central server can calculate the third weight update parameter based on the first weight update parameter and the second weight update parameter obtained from other training devices. The third weight update parameters sent by the central server are received to adjust the model to be trained, resulting in the trained inference model.

[0069] like Figure 2 As shown, the local model can be installed on the local model training node of Institution A (medical institution) or Institution B (psychological platform), and the raw data is provided by the local anonymized database for training. The central server is set up in the federated server of TEE (Trusted Execution Environment).

[0070] Specifically, the central server uses a federated averaging algorithm to comprehensively calculate the first and second weight update parameters. Other training devices refer to remote training devices equipped with the same model as the local model, and there can be multiple such devices. This not only solves the problem of data not leaving the local domain, but also, due to global parallel aggregation, the training effect is equivalent to training with the data centrally located.

[0071] In this federated learning framework, the local model is an initial "empty shell AI model" distributed by the central server, which is trained locally by each institution. Institutions A and B only upload encrypted gradient weight parameters, thus achieving distributed AI training where "data is available but not visible, and the model moves even when the data doesn't." The central server provides updated third-weight parameters to each local model by distributing global parameters.

[0072] In one embodiment, before the central server updates the parameters based on the first weight and the second weight obtained from other training devices, the first weight update parameters can be split based on a secure multi-party computation protocol to obtain a first parameter ciphertext, so that the central server can calculate the third weight update parameters based on the first parameter ciphertext and the second parameter ciphertext obtained from other training devices based on the split second weight update parameters.

[0073] Splitting based on a secure multi-party computation protocol refers to using cryptographic protocols (Secret Sharing) or homomorphic encryption technology to split the data of multiple institutions into scrambled fragments when jointly computing a model parameter, thus obtaining the ciphertext of the first parameter.

[0074] The second parameter ciphertext is similar to the first parameter ciphertext, except that the data source is the second weight update parameter.

[0075] The central server performs addition and multiplication on the first and second parameter ciphertexts in encrypted form, finally piecing together the correct result. This process utilizes the Secure Multi-Party Computation (MPC) algorithm, which avoids using the original plaintext data while ensuring the calculation result is identical to that using the original plaintext data. The globally shared model update module distributes global parameters, which are the third weight update parameters.

[0076] The cross-agency collaboration solutions provided by the aforementioned federal training reduce technical barriers to data sharing and provide infrastructure for collaborative research.

[0077] In one embodiment, a call log for the first training data can be generated, and the hash digest of the call log can be stored on a blockchain. If there are multiple institutions, call logs can be generated in the same way for the training process of each institution. In another embodiment, cross-institutional access permissions can be set. For example, since the original data is not shared, the access is actually to "local computing power." For instance, university lab A initiates a request: "I want to run a large model training task on institution B's server." After institution B's administrator approves the request, the authorization service issues a temporary digital key, allowing the lab's training script to legally access institution B's local federated node.

[0078] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0079] The artificial intelligence-based psychological assessment and medical auxiliary diagnosis solutions involved in the embodiments of the present invention strictly comply with national regulations on data security, personal information protection and medical industry management in terms of data collection, processing, model training and application, and meet the requirements of medical ethics review.

[0080] The diagnostic data involved in the embodiments of this invention (such as, but not limited to, behavioral observation data, clinical consultation records, physiological and biochemical indicators, etc.) are all derived from legal and compliant publicly available de-identified datasets, or legally collected with the subject's (or their legal guardian's) explicit consent after full disclosure. During the data collection process, this invention does not involve concealment, deception, or unauthorized acquisition of sensitive personal information.

[0081] In the embodiments of this invention, when collecting personal biometric information such as facial micro-expressions, eye gaze trajectories, voice acoustic features, or body movements of subjects for auxiliary assessment in psychological-related diagnoses, the collection is conducted in a specific environment where the subjects are fully informed and voluntarily cooperate. The subjects have been clearly informed of the specific purpose and scope of use of the data collection and have obtained their specific consent. This invention does not contain any covert, non-intrusive, or unauthorized image and sound capture methods.

[0082] If the embodiments of this invention involve multimodal medical data fusion (e.g., simultaneously inputting text questionnaire data, facial expression videos, and voice recordings of subjects), this system adopts a strict permission isolation and encrypted transmission mechanism when processing such data. The original data of different modalities are physically or logically isolated during the feature extraction stage. The artificial intelligence algorithm model only performs fusion calculations on the extracted abstract feature vectors. The entire calculation process and the final output diagnostic results cannot be reversed to reconstruct the original multimodal joint information of the subject, thus eliminating the privacy leakage risk that may be caused by cross-modal data association.

[0083] Before using the aforementioned data for model training or actual diagnosis, all data underwent rigorous de-identification and anonymization processes. This eliminated identity-related biases while retaining the disease characteristic dimensions required for psychological assessment. Furthermore, the algorithmic decision-making mechanism of this invention is built upon objective psychopathological characteristics. During model training, it fully considers the diversity, representativeness, and balance of the samples, and avoids setting biased, stereotypical, or discriminatory decision rules based on non-disease factors such as the subject's gender, age, occupation, or social status.

[0084] The systems, methods, and devices claimed in this invention for psychological assessment and medical auxiliary diagnosis, etc., whose output analysis results, risk probabilities, or classification labels are intended solely for clinical reference by legally qualified psychiatrists, psychotherapists, and other medical professionals. This invention serves only as a clinical auxiliary assessment tool and cannot be used directly as the final basis for clinical diagnosis, nor can it automatically generate drug treatment prescriptions or psychological intervention instructions. The final psychological state assessment and treatment plan must be made by a professional physician through interviews and comprehensive clinical examinations. The implementation of this invention does not replace the diagnostic and treatment actions of professional medical personnel.

[0085] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for enhancing privacy and managing compliance of highly sensitive mental health data, characterized in that, include: Obtain raw data from psychological counseling scenarios; For the data scenario of the original data, the original data is processed using the privacy enhancement method corresponding to the data scenario to obtain the first training data; The first training data is used to train the model to be trained, and the first weight update parameters of the model to be trained are obtained based on the DP-SGD algorithm. The model to be trained is adjusted based on the first weight update parameter to obtain the trained inference model. Input data is provided to the trained inference model to obtain the inference result of the inference model. If abnormal information is detected in the inference result, the inference result is intercepted. The input data is multimodal medical data.

2. The privacy enhancement and compliance management method for highly sensitive mental health data as described in claim 1, characterized in that, The method further includes: A five-level sensitivity classification model is used to label the sensitivity level and data category of the original data.

3. The privacy enhancement and compliance management method for highly sensitive mental health data as described in claim 2, characterized in that, The method further includes: Generate an interactive electronic authorization contract that matches the sensitivity level of the original data; The interactive electronic authorization contract is encrypted and stored using a blockchain.

4. The privacy enhancement and compliance management method for highly sensitive mental health data as described in claim 1, characterized in that, The privacy enhancement methods include: Perform a hash operation on the original data; Alternatively, the original data can be replaced using a compliant generalization algorithm based on K-anonymity; Alternatively, noisy data can be added to the original data based on differential privacy technology.

5. The privacy enhancement and compliance management method for highly sensitive mental health data as described in claim 1, characterized in that, Before training the model to be trained using the first training data, the method further includes: A re-identification attack simulation is performed on the first training data. If the simulation result indicates that the de-identification of the first training data has failed, the privacy enhancement processing is performed on the first training data again. And / or, perform semantic availability verification on the first training data. If the verification result indicates that the information entropy loss of the first training data is too high, then perform noise reduction processing on the first training data.

6. The privacy enhancement and compliance management method for highly sensitive mental health data as described in claim 1, characterized in that, The model to be trained is a local model; the step of adjusting the model to be trained based on the first weight update parameters to obtain the trained inference model includes: The first weight update parameter is sent to the central server so that the central server can calculate the third weight update parameter based on the first weight update parameter and the second weight update parameter obtained from other training devices. The training model is adjusted by receiving the third weight update parameters sent by the central server to obtain the trained inference model.

7. The privacy enhancement and compliance management method for highly sensitive mental health data as described in claim 6, characterized in that, Before the central server updates the parameters according to the first weight and the second weight obtained from other training devices, the method further includes: The first weight update parameter is split based on a secure multi-party computation protocol to obtain a first parameter ciphertext, so that the central server can calculate the third weight update parameter based on the first parameter ciphertext and the second parameter ciphertext obtained from other training devices based on the split second weight update parameter.

8. The privacy enhancement and compliance management method for highly sensitive mental health data as described in claim 1, characterized in that, The method further includes: A call log for the first training data is generated, and the hash digest of the call log is stored on the blockchain.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the privacy enhancement and compliance management method for highly sensitive mental health data as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when executed by a processor, implements the privacy enhancement and compliance management method for highly sensitive mental health data as described in any one of claims 1-8.

Images