Intelligent monitoring method and system for digital supervision platform

Abstract

The invention belongs to the technical field of digital monitoring, and provides an intelligent monitoring method and system for a digital supervision platform, and the method comprises the steps: carrying out the sensitivity grading of data, the homomorphic encryption of high-sensitivity data, and the secure multi-party calculation, achieving the secret state feature alignment, and generating a zero-knowledge proof; selecting a federal mode according to data sensitivity, injecting adaptive differential privacy noise, and encrypting transmission gradient through two channels; the aggregation gradient of the Byzantine fault-tolerant protocol is improved, abnormal nodes are detected, and reputation punishment is carried out; constructing a collaborative decision tree based on credible data, generating zero-knowledge proof and performing uplink storage, and supporting non-intrusive auditing; establishing a feedback mechanism, and dynamically adjusting privacy budget to balance privacy and monitoring efficiency; according to the method, the problems of cross-mechanism data collaboration privacy protection and efficiency balance are solved, secret state calculation, dynamic privacy adjustment and whole-process credible evidence storage are realized, and the monitoring safety and the supervision efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the field of digital monitoring technology, and in particular relates to an intelligent monitoring method and system for a digital supervision platform. Background Art

[0002] With the rapid development of digital technology, digital supervision platforms are increasingly being used in finance, healthcare, government affairs and other fields. These platforms collect, analyze and make decisions based on multi-source monitoring data, enabling real-time supervision and risk control for various business scenarios. However, data privacy and security issues are becoming increasingly prominent during the monitoring process, primarily in the following areas:

[0003] While traditional encryption algorithms can protect data transmission security, they cannot support secret computing, resulting in data leakage risks even after decryption. Frequent encryption and decryption operations also significantly reduce computing efficiency.

[0004] Existing federated learning frameworks lack self-optimization mechanisms in terms of dynamic adaptation to data sensitivity and privacy budget allocation, making it difficult to balance model accuracy and privacy protection in highly sensitive data scenarios.

[0005] In distributed monitoring scenarios, traditional Byzantine fault-tolerant protocols (such as PBFT) have high communication overhead when performing dense gradient aggregation, making them unable to adapt to the real-time collaboration needs of large-scale nodes.

[0006] When the data feature spaces of different institutions are heterogeneous, the lack of secure feature alignment technology makes cross-institutional model training ineffective;

[0007] The non-traceability and non-verifiability of the monitoring process make it difficult for regulators to conduct compliance audits without accessing the original data, failing to meet the need for non-intrusive supervision.

[0008] To this end, the present invention provides an intelligent monitoring method and system for a digital supervision platform. Summary of the Invention

[0009] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0010] The technical solution adopted by the present invention to solve its technical problem is:

[0011] In a first aspect, the present invention provides an intelligent monitoring method for a digital supervision platform, comprising:

[0012] S1: Local monitoring data is graded according to its sensitivity, highly sensitive data is processed using homomorphic encryption, cross-institutional confidential feature alignment is achieved through secure multi-party computation, feature hashing rules are negotiated based on smart contracts, and zero-knowledge proofs are generated.

[0013] S2: Dynamically select the federation mode based on data sensitivity, inject adaptive differential privacy noise into the local monitoring model, and use a dual-channel encryption algorithm, homomorphic encryption and symmetric encryption, to encrypt and transmit the monitoring model gradient.

[0014] S3: Aggregate the gradients of the encrypted transmission monitoring model through an improved Byzantine fault-tolerant protocol, detect abnormal nodes, and trigger a reputation penalty mechanism to filter and obtain trusted data.

[0015] S4: Based on the obtained trusted data, a multi-party collaborative decision tree model is constructed to generate monitoring conclusions. Using zero-knowledge proof technology, a verifiable zero-knowledge proof is generated for the monitoring process. The proof hash and key operation logs are stored on-chain, supporting regulators to conduct non-invasive audits through smart contracts.

[0016] S5: Establish a feedback evaluation mechanism to dynamically adjust the privacy budget based on monitoring accuracy and privacy leakage risks, and continuously optimize the balance between privacy security and efficient monitoring.

[0017] As a further improvement of the present invention, the specific process of performing sensitivity classification on the local monitoring data is as follows:

[0018] First, the local monitoring data is subjected to noise filtering, missing value filling, and numerical feature standardization. Based on the local monitoring data after noise filtering, missing value filling, and numerical feature standardization, sensitivity classification is performed. The local monitoring data is divided into two levels according to the type of local monitoring data and privacy risks: high-sensitivity data and low-sensitivity data.

[0019] Data that can directly identify personal information is set as high-sensitivity data, and data that cannot directly identify personal information and local monitoring data that has been made public are set as low-sensitivity data.

[0020] As a further improvement of the present invention, the specific process of using homomorphic encryption to process highly sensitive data is as follows:

[0021] The Paillier additive homomorphic encryption algorithm is used to convert highly sensitive data into a secret state, converting it into data that supports addition operations under ciphertext. The Paillier algorithm is used to generate a public-private key pair, and the public key is used to encrypt the highly sensitive data. The encrypted highly sensitive data can only be decrypted using the private key.

[0022] As a further improvement of the present invention, the specific process of achieving cross-organization secret feature alignment through secure multi-party computation is as follows:

[0023] Each participant generates a feature hash value locally and compares the hash consistency through the MPC protocol without exposing the original features;

[0024] Create hash indexes for shared features, mark non-shared features as private features, and only store hash values ​​locally;

[0025] By using the private set intersection (PSI) technology of MPC, the local monitoring data that needs to be aligned across institutions is screened out through the common sample ID determined by PSI.

[0026] As a further improvement of the present invention, the process of negotiating the feature hash rule based on the smart contract and generating the zero-knowledge proof is as follows:

[0027] Predefine feature hashing rules. Each institution generates feature hashes based on these rules to ensure a unified hash format across institutions. Define feature hashing rules as the public input for zero-knowledge proofs. The prover automatically triggers the zk-SNARK generation function based on these rules to generate proof π, which automatically triggers the zero-knowledge proof generation function to verify the legitimacy of the feature alignment results, proving that the local feature hash is consistent with the cross-institutional alignment results without leaking the original feature values.

[0028] The verifier calls the verification function through the smart contract and uses rule verification to confirm that the feature alignment meets the preset hash rules;

[0029] After verification, the alignment results and the proof hash are stored on the chain as the original basis for the audit.

[0030] As a further improvement of the present invention, the specific process of dynamically selecting the federation mode according to data sensitivity is as follows:

[0031] Based on the local monitoring data after sensitivity level classification, the appropriate federated learning mode is dynamically selected. The vertical federated learning mode is adopted for high-sensitivity data, and secure multi-party computing plus Paillier homomorphic encryption is used to perform dense aggregation of the gradients of heterogeneous features. The horizontal federated learning mode is adopted for low-sensitivity data, and the federated averaging method is used to directly aggregate the gradients of homogeneous features.

[0032] As a further improvement of the present invention: the specific process of injecting adaptive differential privacy noise into the local monitoring model is as follows:

[0033] During the training process of the local monitoring model, adaptive differential privacy noise is injected into the model parameters or gradients. The intensity of the adaptive differential privacy noise is dynamically adjusted according to the data sensitivity: high-intensity noise is injected into high-sensitivity data to enhance privacy protection, while low-intensity noise is injected into low-sensitivity data.

[0034] As a further improvement of the present invention, the specific process of using the homomorphic encryption algorithm and the symmetric encryption algorithm to encrypt and transmit the monitoring model gradient is as follows:

[0035] In the vertical federated learning mode, homomorphic encryption is performed on the computing layer to ensure that highly sensitive data will not be leaked during the calculation process. The symmetric encryption algorithm encrypts the homomorphic ciphertext twice, and the key is exchanged through Diffie-Hellman to ensure that highly sensitive data will not be leaked during data transmission.

[0036] As a further improvement of the present invention, the specific process of abnormal node detection is as follows:

[0037] Calculate the mean and standard deviation of the local monitoring model gradient using the formula: Make a judgment, among which, is the mean of the local monitoring model gradient, is the standard deviation of the local monitoring model gradient, Model gradients for local monitoring;

[0038] when , the node is judged to be an abnormal node.

[0039] In a second aspect, the present invention provides an intelligent monitoring system for a digital supervision platform, comprising:

[0040] Data Hierarchical Encryption Module: This module classifies the sensitivity of local monitoring data, uses homomorphic encryption for highly sensitive data, achieves cross-institutional confidential feature alignment through secure multi-party computing, negotiates feature hashing rules based on smart contracts, and generates zero-knowledge proofs.

[0041] Federated learning scheduling module: Dynamically selects the federation mode based on data sensitivity, injects adaptive differential privacy noise into the local monitoring model, and uses a dual-channel encryption algorithm, homomorphic encryption and symmetric encryption, to transmit the monitoring model gradient.

[0042] Consensus aggregation module: Aggregates the gradients of the encrypted transmission monitoring model through an improved Byzantine fault-tolerant protocol, detects abnormal nodes, and triggers a reputation penalty mechanism to filter and obtain trusted data.

[0043] Model building and audit module: Based on the obtained trusted data, a multi-party collaborative decision tree model is constructed to generate monitoring conclusions. Utilizing zero-knowledge proof technology, a verifiable zero-knowledge proof is generated for the monitoring process. The proof hash and key operation logs are stored on-chain, supporting regulators to conduct non-invasive audits through smart contracts.

[0044] Feedback Optimization Module: Establish a feedback evaluation mechanism to dynamically adjust the privacy budget based on monitoring accuracy and privacy leakage risks, and continuously optimize the balance between privacy security and efficient monitoring.

[0045] The beneficial effects of the present invention are as follows:

[0046] 1. Through homomorphic encryption, secure multi-party computation, and zero-knowledge proof, highly sensitive data is ensured to be processed confidentially throughout cross-institutional collaboration, preventing the leakage of original data and meeting the privacy compliance requirements of scenarios such as healthcare and finance. The federation mode (vertical / horizontal) is dynamically selected based on data sensitivity. In combination with adaptive differential privacy noise injection and dynamic privacy budget adjustment, privacy is protected in highly sensitive scenarios while improving model accuracy in less sensitive scenarios. The Byzantine fault-tolerant protocol is improved to achieve confidential gradient aggregation. Abnormal node detection and a credit penalty mechanism are used to resist Byzantine attacks and ensure the credibility of aggregation results.

[0047] 2. Utilizing zero-knowledge proof and blockchain evidence storage, key operation logs and proof hashes of the monitoring process are uploaded to the chain, supporting non-invasive audits by regulators, ensuring process compliance and traceability, and improving audit efficiency. The feedback and evaluation mechanism dynamically adjusts the privacy budget based on reinforcement learning, combining real-time optimization with model accuracy and privacy risks to form a monitoring-assessment-optimization closed loop, continuously improving system security and monitoring efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention will be further described below with reference to the accompanying drawings.

[0049] Figure 1 is a flow chart of the steps of the present invention;

[0050] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0051] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0052] Example 1

[0053] like Figure 1 As shown, an intelligent monitoring method for a digital supervision platform according to an embodiment of the present invention includes:

[0054] S1: Local monitoring data is graded according to its sensitivity, highly sensitive data is processed using homomorphic encryption, cross-institutional confidential feature alignment is achieved through secure multi-party computation, feature hashing rules are negotiated based on smart contracts, and zero-knowledge proofs are generated.

[0055] First, the local monitoring data is subjected to noise filtering, missing value filling, and numerical feature standardization;

[0056] It should be noted that local monitoring data includes but is not limited to structured information including transaction amounts, medical records, patient personal information, and electronic medical records;

[0057] For example, the transaction amount feature is standardized to make the data distribution uniform with a mean of 0 and a standard deviation of 1, which facilitates subsequent cross-institutional feature alignment;

[0058] Sensitivity grading is performed based on local monitoring data after noise filtering, missing value filling, and numerical feature standardization;

[0059] The specific process of sensitivity grading is as follows:

[0060] According to the type of local monitoring data and privacy risks, local monitoring data is divided into two levels: high-sensitivity data and low-sensitivity data;

[0061] For example, in a medical monitoring scenario, data that can directly identify personal information, including medical records, patient personal information, electronic medical records, and medical diagnosis results, is set as highly sensitive data;

[0062] Set data that cannot directly identify personal information, such as patient age, gender, and length of hospital stay, and publicly available local monitoring data, as low-sensitivity data;

[0063] The specific process of using homomorphic encryption for highly sensitive data is as follows:

[0064] Use the Paillier additive homomorphic encryption algorithm to convert highly sensitive data into a secret state;

[0065] For example, the patient's blood pressure Get the ciphertext , ciphertext supports addition operations under ciphertext, for example, , ensuring that cross-institutional calculations do not require decryption of original data;

[0066] When performing cross-institutional calculations, local monitoring data must be calculated without decryption to protect privacy.

[0067] Generate public and private key pairs using the Paillier algorithm , public key Distribute to all participants and use public keys to access highly sensitive data encryption: ,in, , , is a large prime number, is a random number;

[0068] Encrypted highly sensitive data can only be accessed through the private key Decryption to ensure privacy and security during transmission and calculation;

[0069] The specific process of achieving cross-institutional secret feature alignment through secure multi-party computing is as follows:

[0070] Each participant generates a feature hash value locally and compares the hash consistency through the MPC protocol without exposing the original features;

[0071] Create hash indexes for shared features, mark non-shared features as private features, and only store hash values ​​locally;

[0072] Using MPC's private set intersection (PSI) technology, the shared sample IDs determined by PSI are used to filter out local monitoring data that need to be aligned across institutions, preventing irrelevant samples from participating in calculations and improving efficiency and privacy protection.

[0073] The specific process of negotiating feature hashing rules and generating zero-knowledge proof based on smart contracts:

[0074] Smart contracts predefine feature hashing rules, and each institution generates feature hashes based on these rules, ensuring a unified hash format across institutions and supporting consistency verification of subsequent zero-knowledge proofs.

[0075] Define feature hashing rules as the public input of zero-knowledge proof. The prover automatically triggers the zk-SNARK generation function based on the rules to generate proof π, which automatically triggers the zero-knowledge proof generation function to verify the legitimacy of the feature alignment results, proving that the local feature hash is consistent with the cross-institutional alignment result and does not leak the original feature value.

[0076] The verifier calls the verification function through the smart contract ,Use rule verification to confirm that the feature alignment complies with the preset hash rules;

[0077] After verification, the alignment results and the proof hash are stored on the chain as the original basis for auditing;

[0078] S2: Dynamically selects the federation mode based on the sensitivity data, injects adaptive differential privacy noise into the local monitoring model, and uses a dual-channel encryption algorithm, homomorphic encryption and symmetric encryption, to transmit the local monitoring model gradient.

[0079] The specific process of dynamically selecting the federation mode based on data sensitivity is as follows:

[0080] Based on the local monitoring data divided into sensitivity levels, the appropriate federated learning model is dynamically selected. High-sensitivity data adopts the vertical federated learning model, which uses secure multi-party computing plus Paillier homomorphic encryption to perform encrypted aggregation of the gradients of heterogeneous features to avoid cross-institutional transmission of raw data. Low-sensitivity data can adopt the horizontal federated learning model, which uses the federated averaging method to directly aggregate the gradients of homogeneous features to improve training efficiency.

[0081] The specific process of injecting adaptive differential privacy noise into the local monitoring model is as follows:

[0082] During the training of the local monitoring model, adaptive differential privacy noise is injected into the model parameters or gradients. The intensity of the adaptive differential privacy noise is dynamically adjusted according to the data sensitivity: more noise is injected into high-sensitivity data to enhance privacy protection, while weaker noise is injected into low-sensitivity data to reduce the impact on model accuracy.

[0083] It should be noted that in the vertical federated learning mode, high-intensity noise is injected to ensure the privacy and security of individual data. In the horizontal federated learning mode, the intensity of noise injection is reduced to reduce the accuracy loss of the local monitoring model.

[0084] When the validation set accuracy of the local monitoring model is less than or equal to 85%, the noise injection intensity is increased to ensure the privacy and security of individual data;

[0085] When the validation set accuracy of the local monitoring model is greater than 85%, the noise injection intensity is reduced to ensure the accuracy of the local monitoring model;

[0086] This mechanism achieves a balance between privacy protection and model performance by dynamically adjusting the privacy budget PrivacyBudget;

[0087] The specific process of using the homomorphic encryption algorithm and the symmetric encryption algorithm to encrypt and transmit the local monitoring model gradient through dual channels is as follows:

[0088] Homomorphic encryption algorithms ensure privacy and security during the aggregation process. In the vertical federated learning mode, homomorphic encryption is performed on the computing layer to prevent the leakage of highly sensitive data during the calculation process. Symmetric encryption algorithms encrypt homomorphic ciphertext twice, and the key is exchanged through Diffie-Hellman to ensure that highly sensitive data is not leaked during data transmission.

[0089] Homomorphic encryption: allows calculations on data in ciphertext, ensuring that gradients remain encrypted during transmission and aggregation, allowing model updates to be completed without decryption.

[0090] Symmetric encryption: Use symmetric keys to further encrypt homomorphically encrypted gradients to improve the security of the transmission channel;

[0091] A dual-channel encryption mechanism ensures the confidentiality of local monitoring model gradients during cross-institutional transmission, preventing data leakage or tampering.

[0092] When the local monitoring model gradient is transmitted to the aggregation node, a dual-channel encryption mechanism is used to ensure the privacy and security of the local monitoring model gradient during cross-institutional transmission;

[0093] S3: Aggregate the gradients of the encrypted transmission local monitoring model through an improved Byzantine fault-tolerant protocol, detect abnormal nodes, and trigger a reputation penalty mechanism to filter and obtain trusted data;

[0094] The specific process of aggregating the encrypted transmission local monitoring model gradient through the improved Byzantine fault tolerance protocol is as follows:

[0095] The improved BFT protocol supports encrypted aggregation of homomorphically encrypted model gradients. Leveraging the additive homomorphic nature of homomorphic encryption, each node can sum the ciphertext gradients without decrypting them, ensuring that data remains encrypted during the aggregation process.

[0096] Define the additive aggregation formula of the ciphertext gradient: , calculate the aggregation result, where It is the local monitoring model gradient encrypted by the dual-channel encryption mechanism. , and is a large prime number; it realizes gradient aggregation in a secret state without decrypting the original local monitoring model gradient, ensuring the privacy of the aggregation process;

[0097] By optimizing the BFT consensus process, the communication overhead of cross-institutional gradient aggregation is reduced, adapting to large-scale distributed monitoring scenarios;

[0098] The process of abnormal node detection is as follows:

[0099] Compare the local monitoring model gradient uploaded by each node with the aggregation results to determine abnormal nodes;

[0100] Specifically, calculate the mean and standard deviation of the local monitoring model gradient using the formula: Make a judgment, among which, is the mean of the local monitoring model gradient, is the standard deviation of the local monitoring model gradient, Model gradients for local monitoring;

[0101] when When , the node is judged to be an abnormal node;

[0102] Reputation penalty mechanism and trusted node screening process:

[0103] Each node initially has a basic credit score. Each time it participates in gradient aggregation, the credit score is adjusted based on its gradient consistency and response speed indicators.

[0104] Abnormal nodes will have their reputation points deducted, and when the score falls below the threshold, a penalty mechanism will be triggered;

[0105] For example, temporarily isolate and reduce the aggregation weight, and automatically execute reputation score updates and penalty measures based on preset rules through smart contracts; for example, when a node is detected as abnormal three times in a row, the smart contract automatically removes it from the trusted node list until it passes re-verification;

[0106] S4: Based on the obtained trusted data, a multi-party collaborative decision tree model is constructed to generate monitoring conclusions. Using zero-knowledge proof technology, a verifiable zero-knowledge proof is generated for the monitoring process. The proof hash and key operation logs are stored on-chain, supporting regulators to conduct non-invasive audits through smart contracts.

[0107] The specific process of building a multi-party collaborative decision tree model based on the obtained credible data to generate monitoring conclusions is as follows:

[0108] Based on the obtained trusted data, the participating institutions jointly built a decision tree model, using a horizontal federated learning model. Each node trained a subtree locally, encrypted the parameters and aggregated them to ensure data reliability and avoid the centralized storage of the original data.

[0109] Select the vertical federated learning model. When the data feature spaces of each institution are different, secure multi-party computing (MPC) is used to jointly calculate the feature importance in an encrypted state to collaboratively build a cross-institutional decision tree.

[0110] Only use encrypted gradients or aggregated results provided by screened high-reputation nodes to ensure that decision trees are built based on reliable data and prevent malicious or abnormal data from affecting the accuracy of conclusions.

[0111] The specific process of generating and verifying zero-knowledge proof:

[0112] Using zero-knowledge proof technology, a verifiable zero-knowledge proof is generated for the entire monitoring process; the proof content includes:

[0113] In S1, the compliance of encryption algorithms;

[0114] In S2, verify the transmission security;

[0115] In S3, ensure data credibility;

[0116] In S4, the decision tree construction logic;

[0117] Zero-knowledge proof only proves the legitimacy of the monitoring process and does not disclose original data, model parameters, or intermediate calculation results, ensuring the unity of privacy protection and process credibility;

[0118] The specific process of storing the proof hash and key operation log on the chain:

[0119] Store the hash value of the zero-knowledge proof and key operation logs in the blockchain;

[0120] Hash storage ensures that data cannot be tampered with, logging provides traceability, and the distributed nature of blockchain ensures the security and reliability of storage;

[0121] Regulators can verify compliance of the monitoring process by using smart contracts to call proof hashes and logs on the blockchain without having to access the original monitoring system or decrypt the data:

[0122] Smart contracts automatically verify the validity of zero-knowledge proofs, confirm whether the monitoring process complies with preset rules, compare the consistency of operation logs with the proof content, and check whether there is process tampering or abnormal operation. Non-intrusive audits avoid interference with normal monitoring operations while ensuring transparency and efficiency of supervision.

[0123] S5: Establish a feedback evaluation mechanism to dynamically adjust the privacy budget based on monitoring accuracy and privacy leakage risks, and continuously optimize the balance between privacy security and efficient monitoring.

[0124] Specific process of establishing feedback evaluation mechanism:

[0125] Based on the output results of the multi-party collaborative decision tree model, the accuracy, recall rate, and F1 score indicators are calculated by comparing them with the actual labels of the actual monitoring targets;

[0126] First define the threshold range:

[0127] For highly sensitive data: the threshold is defined as precision ≥ 80% and recall ≥ 75%;

[0128] For low-sensitivity data: the threshold is defined as accuracy ≥ 95%, F1 ≥ 0.95;

[0129] The model accuracy requirements in highly sensitive data monitoring scenarios can be appropriately lowered, while lower sensitivity scenarios require higher accuracy.

[0130] The current privacy protection level is evaluated by the remaining amount of the privacy budget. The smaller the privacy budget, the stronger the privacy protection.

[0131] Build a privacy leakage risk model based on historical data, analyze the potential attack surface, combine the current encryption mechanism and noise injection strategy, and output the risk probability value;

[0132] The process of dynamically adjusting the privacy budget is:

[0133] Using reinforcement learning algorithms, the privacy budget allocation of each link is dynamically adjusted based on real-time evaluation results; for example:

[0134] When the monitoring accuracy is lower than the threshold, the differential privacy noise intensity is reduced to increase the privacy budget value, releasing part of the privacy budget to improve the model accuracy;

[0135] When the privacy leakage risk assessment value is higher than the threshold, the noise intensity is increased to reduce the privacy budget value, consuming more privacy budget to enhance protection;

[0136] The technical solution of the embodiment of the present invention is: data sensitivity classification, homomorphic encryption of highly sensitive data, secure multi-party computing to achieve secret feature alignment and generate zero-knowledge proof; select the federation mode according to data sensitivity, inject adaptive differential privacy noise, and dual-channel encrypted transmission gradient; improve the Byzantine fault tolerance protocol to aggregate gradients, detect abnormal nodes and impose reputation penalties; build a collaborative decision tree based on trusted data, generate zero-knowledge proof and store it on the chain, and support non-invasive auditing; establish a feedback mechanism to dynamically adjust the privacy budget to balance privacy and monitoring efficiency.

[0137] Example 2

[0138] like Figure 2 As shown, based on Example 1, the present invention provides a system for intelligent monitoring of a digital supervision platform, comprising:

[0139] Data Hierarchical Encryption Module: This module classifies the sensitivity of local monitoring data, uses homomorphic encryption for highly sensitive data, achieves cross-institutional confidential feature alignment through secure multi-party computing, negotiates feature hashing rules based on smart contracts, and generates zero-knowledge proofs.

[0140] Federated learning scheduling module: Dynamically selects the federation mode based on data sensitivity, injects adaptive differential privacy noise into the local monitoring model, and uses a dual-channel encryption algorithm, homomorphic encryption and symmetric encryption, to transmit the monitoring model gradient.

[0141] Consensus aggregation module: Aggregates the gradients of the encrypted transmission monitoring model through an improved Byzantine fault-tolerant protocol, detects abnormal nodes, and triggers a reputation penalty mechanism to filter and obtain trusted data.

[0142] Model building and audit module: Based on the obtained trusted data, a multi-party collaborative decision tree model is constructed to generate monitoring conclusions. Utilizing zero-knowledge proof technology, a verifiable zero-knowledge proof is generated for the monitoring process. The proof hash and key operation logs are stored on-chain, supporting regulators to conduct non-invasive audits through smart contracts.

[0143] Feedback Optimization Module: Establish a feedback evaluation mechanism to dynamically adjust the privacy budget based on monitoring accuracy and privacy leakage risks, and continuously optimize the balance between privacy security and efficient monitoring.

[0144] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart monitoring method for a digital supervision platform, characterized by: include: S1: Local monitoring data is graded according to its sensitivity, highly sensitive data is processed using homomorphic encryption, cross-institutional confidential feature alignment is achieved through secure multi-party computation, feature hashing rules are negotiated based on smart contracts, and zero-knowledge proofs are generated. S2: Dynamically select the federation mode based on data sensitivity, inject adaptive differential privacy noise into the local monitoring model, and use a dual-channel encryption algorithm, homomorphic encryption and symmetric encryption, to encrypt and transmit the monitoring model gradient. S3: Aggregate the gradients of the encrypted transmission monitoring model through an improved Byzantine fault-tolerant protocol, detect abnormal nodes, and trigger a reputation penalty mechanism to filter and obtain trusted data. S4: Based on the obtained trusted data, a multi-party collaborative decision tree model is constructed to generate monitoring conclusions. Using zero-knowledge proof technology, a verifiable zero-knowledge proof is generated for the monitoring process. The proof hash and key operation logs are stored on-chain, supporting regulators to conduct non-invasive audits through smart contracts. S5: Establish a feedback evaluation mechanism to dynamically adjust the privacy budget based on monitoring accuracy and privacy leakage risks, and continuously optimize the balance between privacy security and efficient monitoring.

2. The intelligent monitoring method for a digital supervision platform according to claim 1, characterized in that: The specific process of sensitivity classification of local monitoring data is as follows: First, the local monitoring data is subjected to noise filtering, missing value filling, and numerical feature standardization. Based on the local monitoring data after noise filtering, missing value filling, and numerical feature standardization, sensitivity classification is performed. The local monitoring data is divided into two levels according to the type of local monitoring data and privacy risks: high-sensitivity data and low-sensitivity data. Data that can directly identify personal information is set as high-sensitivity data, and data that cannot directly identify personal information and local monitoring data that has been made public are set as low-sensitivity data.

3. The intelligent monitoring method and system for a digital supervision platform according to claim 1, characterized in that: The specific process of using homomorphic encryption for highly sensitive data is as follows: The Paillier additive homomorphic encryption algorithm is used to convert highly sensitive data into a secret state, converting it into data that supports addition operations under ciphertext. The Paillier algorithm is used to generate a public-private key pair, and the public key is used to encrypt the highly sensitive data. The encrypted highly sensitive data can only be decrypted using the private key.

4. The intelligent monitoring method for a digital supervision platform according to claim 1, characterized in that: The specific process of achieving cross-institutional secret feature alignment through secure multi-party computation is as follows: Each participant generates a feature hash value locally and compares the hash consistency through the MPC protocol without exposing the original features; Create hash indexes for shared features, mark non-shared features as private features, and only store hash values ​​locally; By using the private set intersection (PSI) technology of MPC, the local monitoring data that needs to be aligned across institutions is screened out through the common sample ID determined by PSI.

5. The intelligent monitoring method for a digital supervision platform according to claim 1, characterized in that: The process of negotiating feature hash rules and generating zero-knowledge proof based on smart contracts is as follows: Predefine feature hashing rules. Each institution generates feature hashes based on these rules to ensure a unified hash format across institutions. Define feature hashing rules as the public input for zero-knowledge proofs. The prover automatically triggers the zk-SNARK generation function based on these rules to generate proof π, which automatically triggers the zero-knowledge proof generation function to verify the legitimacy of the feature alignment results, proving that the local feature hash is consistent with the cross-institutional alignment results without leaking the original feature values. The verifier calls the verification function through the smart contract and uses rule verification to confirm that the feature alignment meets the preset hash rules; After verification, the alignment results and the proof hash are stored on the chain as the original basis for the audit.

6. The intelligent monitoring method for a digital supervision platform according to claim 1, characterized in that: The specific process of dynamically selecting the federation mode based on data sensitivity is as follows: Based on the local monitoring data after sensitivity level classification, the appropriate federated learning mode is dynamically selected. The vertical federated learning mode is adopted for high-sensitivity data, and secure multi-party computing plus Paillier homomorphic encryption is used to perform dense aggregation of the gradients of heterogeneous features. The horizontal federated learning mode is adopted for low-sensitivity data, and the federated averaging method is used to directly aggregate the gradients of homogeneous features.

7. The intelligent monitoring method for a digital supervision platform according to claim 1, characterized in that: The specific process of injecting adaptive differential privacy noise into the local monitoring model is as follows: During the training process of the local monitoring model, adaptive differential privacy noise is injected into the model parameters or gradients. The intensity of the adaptive differential privacy noise is dynamically adjusted according to the data sensitivity: high-intensity noise is injected into high-sensitivity data to enhance privacy protection, while low-intensity noise is injected into low-sensitivity data.

8. The intelligent monitoring method for a digital supervision platform according to claim 1, characterized in that: The specific process of using the dual-channel encryption of homomorphic encryption algorithm and symmetric encryption algorithm to transmit the monitoring model gradient is as follows: In the vertical federated learning mode, homomorphic encryption is performed on the computing layer to ensure that highly sensitive data will not be leaked during the calculation process. The symmetric encryption algorithm encrypts the homomorphic ciphertext twice, and the key is exchanged through Diffie-Hellman to ensure that highly sensitive data will not be leaked during data transmission.

9. The intelligent monitoring method for a digital supervision platform according to claim 1, characterized in that: The specific process of abnormal node detection is as follows: Calculate the mean and standard deviation of the local monitoring model gradient using the formula: Make a judgment, among which, is the mean of the local monitoring model gradient, is the standard deviation of the local monitoring model gradient, Model gradients for local monitoring; when , the node is judged to be an abnormal node.

10. An intelligent monitoring system for a digital supervision platform, characterized by: include: Data Hierarchical Encryption Module: This module classifies the sensitivity of local monitoring data, uses homomorphic encryption for highly sensitive data, achieves cross-institutional confidential feature alignment through secure multi-party computing, negotiates feature hashing rules based on smart contracts, and generates zero-knowledge proofs. Federated learning scheduling module: Dynamically selects the federation mode based on data sensitivity, injects adaptive differential privacy noise into the local monitoring model, and uses a dual-channel encryption algorithm, homomorphic encryption and symmetric encryption, to transmit the monitoring model gradient. Consensus aggregation module: Aggregates the gradients of the encrypted transmission monitoring model through an improved Byzantine fault-tolerant protocol, detects abnormal nodes, and triggers a reputation penalty mechanism to filter and obtain trusted data. Model building and audit module: Based on the obtained trusted data, a multi-party collaborative decision tree model is constructed to generate monitoring conclusions. Utilizing zero-knowledge proof technology, a verifiable zero-knowledge proof is generated for the monitoring process. The proof hash and key operation logs are stored on-chain, supporting regulators to conduct non-invasive audits through smart contracts. Feedback Optimization Module: Establish a feedback evaluation mechanism to dynamically adjust the privacy budget based on monitoring accuracy and privacy leakage risks, and continuously optimize the balance between privacy security and efficient monitoring.

Images