Privacy protection method based on computing migration and AI behavior trajectory analysis

Abstract

The invention provides a privacy protection method based on computing migration and AI behavior trajectory analysis, and relates to the technical field of privacy protection of the Internet of Things. According to the method, a privacy label is bound to target data, an AI behavior track fingerprint is generated in each AI processing stage, and matching verification is performed on the AI behavior track fingerprint and the privacy label, so that dynamic access control and security audit among terminal equipment, edge nodes and a cloud are realized. On one hand, data can be effectively prevented from being stolen or tampered in the migration process through an encryption transmission and fingerprint verification mechanism; and on the other hand, the fingerprint and label information can be selectively stored in the block chain or the alliance chain side chain, so that the tamper-proofing performance and the distributed credibility of the audit record can be enhanced. The method is suitable for various AI application scenes, the risk of unauthorized calling and sensitive information leakage can be reduced, and the traceability and safety of the whole data processing process are improved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of privacy protection in the Internet of Things, and more specifically, to a privacy protection method based on computational migration and AI behavior trajectory analysis. Background Art

[0002] In the IoT environment, existing privacy protection solutions can usually only provide limited protection for a single scenario or specific data type, and it is difficult to meet the comprehensive security requirements under distributed, multi-node, and multi-model collaborative processing. Especially in the case of widespread application of AI technology, the frequent computing migration and data interaction between terminal devices and cloud / edge nodes can easily lead to unauthorized calls or malicious tampering of user sensitive information, which makes user privacy face greater risks of leakage in IoT scenarios. In addition, traditional access control and auditing methods lack traceability of the entire AI processing process, cannot establish high-precision behavior records for each processing step, and cannot promptly block or trace the source when anomalies occur, resulting in obvious deficiencies in existing technologies in ensuring privacy security in large-scale IoT scenarios.

[0003] In order to solve the above problems, for example, a Chinese patent with authorization announcement number CN105930738B discloses a user privacy protection method for a location-based personalized Internet of Things service system. The core idea is to install a privacy protection client application on the user's smart mobile terminal, and let the user independently configure the application that needs to be location protected. Then, the user's location or movement trajectory is blurred or k-anonymized in combination with a regression analysis model, and the local cache and interception mechanism is used to reduce the chance of the user's real location information being sent to the Internet of Things service provider, thereby protecting the user's sensitive location data to a certain extent and improving the location privacy security in the personalized service process.

[0004] However, the above methods mainly focus on a single type of location-based privacy protection, and are more focused on avoiding the disclosure of user geographic coordinate information by obfuscating or anonymizing user locations and trajectories. They lack a systematic security framework for other types of data involved in AI processing in the IoT environment (such as multi-source perception information collected by devices, cross-node collaborative training data, etc.). Moreover, this method does not provide traceability management for the entire AI processing process, making it difficult to achieve accurate auditing and dynamic control of various operations such as edge node or cloud model reasoning, data forwarding, and model updates. Once an attacker bypasses the protection mechanism outside the client application or tampers with the data before generating k-anonymity, it is difficult for existing solutions to trace and block the entire behavior chain, resulting in the inability to fully meet the comprehensive privacy protection needs of multiple types of data in large-scale IoT scenarios. Therefore, it is necessary to propose a more complete and scalable technical solution to overcome the above shortcomings. Summary of the invention

[0005] In view of the shortcomings of the prior art, the embodiments of the present disclosure provide a privacy protection method based on computational migration and AI behavior trajectory analysis, including: Binding a privacy label to target data in the Internet of Things; wherein the privacy label includes: sensitivity information and usage restriction information of the target data; Before performing AI processing on the target data, determining a processing strategy in an AI behavior space based on the privacy label; wherein the AI ​​behavior space includes: a data preprocessing stage, a model reasoning stage, a result output stage, and a model update stage; During the execution of the AI ​​processing, based on the model version, the calling interface, the execution node, the timestamp, and the information associated with the privacy label of the target data, an AI behavior trajectory fingerprint is generated, and the AI ​​behavior trajectory fingerprint is associated with the privacy label; According to the AI ​​behavior trajectory fingerprint and the privacy label, dynamic access control is performed on the AI ​​processing process; After the AI ​​processing is completed, the generated AI behavior trajectory fingerprint and the corresponding privacy label information are stored in the audit module for verification and tracing.

[0006] As an optional implementation, the AI ​​behavior trajectory fingerprint is generated by a hash algorithm or a digital signature algorithm; The AI ​​behavior trajectory fingerprint includes: model version, calling interface, execution node, timestamp, and key fields associated with the privacy label of the target data.

[0007] As an optional implementation, the dynamic access control includes: In response to the AI ​​behavior trajectory fingerprint matching the privacy label, allowing the AI ​​processing process to continue; In response to the AI ​​behavior trajectory fingerprint not matching the privacy label or detecting out-of-range use, the AI ​​processing process is interrupted or the protection mode is switched, and abnormal information is recorded.

[0008] As an optional implementation, the privacy label includes: The sensitivity level field is used to indicate the confidentiality level of the target data; Usage restriction field, used to describe the AI ​​processing scenario in which the target data is used; The compliance requirement field is used to describe the regulations or industry standards that the target data complies with; The life cycle field is used to indicate the storage or usage period of the target data in the IoT environment.

[0009] As an optional implementation, it also includes: In each stage included in the AI ​​behavior space, the privacy label and the AI ​​behavior trajectory fingerprint are respectively given a dynamic check to restrict the continued use of the target data when the privacy label constraint is not met.

[0010] As an optional implementation, when generating the AI ​​behavior trajectory fingerprint, it also includes: Generating sub-fingerprints for each of the multiple sub-steps in the AI ​​processing process, and merging the sub-fingerprints into the final AI behavior trajectory fingerprint after the processing is completed; The sub-fingerprint is used to distinguish the processing information of different sub-steps during the audit phase.

[0011] As an optional implementation, it also includes: When migrating the target data from the IoT terminal device to the edge node or cloud platform, the privacy label and the AI ​​behavior trajectory fingerprint of the target data are encrypted and transmitted, and after the encrypted information is decrypted and verified at the receiving end, a decision is made as to whether to continue the AI ​​processing process.

[0012] As an optional implementation, the content stored in the audit module includes: The AI ​​behavior trajectory fingerprint corresponding to each AI processing process; Privacy label information associated with the AI ​​behavior trajectory fingerprint; The results of dynamic authorization or access control execution; and Records of detected anomalies or violations; The audit module is deployed on edge nodes or cloud platforms to support verification and violation forensic analysis.

[0013] As an optional implementation, it also includes: storing the association information of the AI ​​behavior trajectory fingerprint and the privacy label in a blockchain or a consortium chain side chain, and verifying and recording the association information through a consensus mechanism.

[0014] As an optional implementation, the Internet of Things environment includes: Multiple resource-constrained terminal devices use the method to bind privacy labels to the collected target data; At least one edge node executes at least a portion of the stages in the AI ​​behavior space and generates the AI ​​behavior trajectory fingerprint.

[0015] Beneficial effects of the present invention: Compared with the existing technology, this application has established a unified privacy label and AI behavior trajectory fingerprint management mechanism. By strongly associating the privacy label with the AI ​​processing process and performing real-time comparison and dynamic access control at each stage, it ensures that user data can comply with the corresponding security policies and compliance restrictions at any processing node or processing link to prevent over-scope use or unauthorized access. In particular, in the process of generating AI behavior trajectory fingerprints, fingerprints are recorded and merged for different sub-steps, which improves the refinement of audit traceability.

[0016] Furthermore, when performing computational migration, the present application combines the encrypted transmission and decryption verification mechanism of the target data and its privacy labels and trajectory fingerprints, which can not only effectively prevent data theft or tampering at the network level or in the cross-node collaboration process, but also can achieve a higher level of privacy protection when transferring data or processing tasks between different nodes or clouds by performing mandatory checks on the decryption and fingerprint matching operations at the receiving end. Compared with traditional methods such as location anonymity and k-anonymity only on the client or a single server, the method disclosed in this application provides privacy label management, dynamic access control, and tamper-proof trajectory auditing functions that can be applied to multiple data types and multiple links, thereby supporting distributed, multi-model collaborative IoT application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a privacy protection method based on computational migration and AI behavior trajectory analysis provided in an embodiment of the present disclosure; Figure 2 A schematic diagram of the components of an AI behavior trajectory provided by an embodiment of the present disclosure; Figure 3 A schematic diagram of the contents of a privacy label provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0019] The terms used in this disclosure are explained here first: AI behavior space: refers to the collection of stages or links that an AI model goes through when processing target data, which may include data preprocessing, model reasoning, result output, model update and other steps.

[0020] AI behavior trajectory fingerprint: refers to the unique identification value generated by a hash algorithm or digital signature algorithm based on the key information (model version, calling interface, execution node, timestamp, information related to privacy labels, etc.) generated during each or each sub-step of the AI ​​processing process. It is used to trace the key elements of the processing process and is tamper-proof.

[0021] Dynamic access control: refers to the process of comparing privacy labels and AI behavior trajectory fingerprints at various stages of AI processing to make real-time decisions on whether to continue or interrupt processing, switch to a more stringent protection mode, etc.

[0022] Out-of-scope use: refers to the nodes, interfaces or functional modules involved in this AI processing exceeding the permitted scope or compliance restrictions specified in the privacy label, such as attempting to upload to an unauthorized cloud platform or calling data fields that are not allowed to be used.

[0023] See also Figure 1 As shown, it is a flowchart of a privacy protection method based on computational migration and AI behavior trajectory analysis provided by an embodiment of the present disclosure, the method includes steps S101 to S105, wherein: S101: Binding a privacy tag to target data in the Internet of Things; wherein the privacy tag includes: sensitivity information and usage restriction information of the target data; S102: Before performing AI processing on the target data, determine a processing strategy in an AI behavior space based on the privacy label; wherein the AI ​​behavior space includes: a data preprocessing stage, a model reasoning stage, a result output stage, and a model update stage; S103: During the execution of the AI ​​processing, based on the model version, the calling interface, the execution node, the timestamp, and the information associated with the privacy label of the target data, an AI behavior trajectory fingerprint is generated, and the AI ​​behavior trajectory fingerprint is associated with the privacy label; S104: Performing dynamic access control on the AI ​​processing process according to the AI ​​behavior trajectory fingerprint and the privacy label; S105: After the AI ​​processing is completed, the generated AI behavior trajectory fingerprint and the corresponding privacy label information are stored in the audit module for verification and tracing.

[0024] In an IoT environment, multiple terminal devices (such as smart home sensors, wearable devices, or industrial monitoring devices) will generate and report various target data. In order to achieve privacy protection based on computing migration and AI behavior trajectory analysis in the IoT environment, for the above S101: In specific implementations, each piece of collected target data can be pre-processed locally on the terminal device or at the initial data collection gateway to identify its potential sensitivity and usable scope. For example, the system can add corresponding privacy labels to the target data based on preset sensitivity classification rules (such as confidential level, general level, etc.) and usage restriction policies (such as only for health monitoring, not for commercial inference, etc.).

[0025] The privacy label includes at least two key fields: "sensitivity information" and "usage restriction information", and is bound to the target data in the form of a data structure or metadata to ensure that the label is consistent with the data during subsequent transmission and calculation processes.

[0026] Regarding S102 above: In the specific implementation, before the target data is prepared for AI processing (such as model reasoning, data mining, real-time analysis, etc.), the pre-defined AI behavior space policy database is queried according to the privacy label carried by the data. The database records the corresponding processing modes for different sensitivities and usage restrictions, such as: only allowing reasoning on local or trusted edge nodes for highly sensitive data; only calling specified AI model versions for data with special usage restrictions, etc.

[0027] After comparing the key information in the privacy label (sensitivity level, usage restrictions, etc.) with the behavioral space strategy, determine the specific strategy and callable model or functional module for this AI processing.

[0028] Exemplarily, the above AI behavior space strategy database may include: Sensitivity level: The confidentiality level of the stored target data, such as HIGH, MEDIUM, LOW, etc. When the system detects the sensitivity level marked in the privacy label, it will query the database to obtain the corresponding processing mode or security measures based on the sensitivity level.

[0029] Usage restriction: records restriction policies for different purposes or scenarios, such as "inference can only be performed on local terminals", "only specified edge nodes are allowed to run", "upload to the public cloud is prohibited", or "only specified versions of AI models are allowed to be called". This field corresponds to the "Usage restriction information" of the privacy label.

[0030] Model version or function module: Indicates which specific version of the AI ​​model or function module the system should call for data of a certain sensitivity or usage restriction. For example, "HIGH + Medical Scenario → Model X_V3.2" can be stored in the database table to select a specific model version for reasoning or training of highly sensitive medical data.

[0031] Compliance requirements: Stores identifiers related to various laws, regulations, or industry standards, such as GDPR, HIPAA, CCPA, or specific region / industry IDs. When the system queries this field, it compares the regulatory information recorded in the privacy label to determine whether the current processing node and environment meet the corresponding compliance conditions. If not, the database will return policy instructions such as "prohibit processing" or "need to switch nodes".

[0032] Available time period or life cycle control: In some application scenarios, there will be time or life cycle restrictions, such as "can only be called from 9:00 to 18:00 on weekdays" and "any reasoning operations are prohibited after data expires". The database can use this field to identify the time period limit or maximum available time for processing operations.

[0033] Regarding S103 above: In specific implementations, when AI processing starts to execute, the system will record the model version involved in this operation (such as the image recognition model of V2.3 or the predictive maintenance model of V1.0), the called interface (such as the API name or service port number), the execution node identifier (such as the edge server ID or cloud node ID), the timestamp, and the key fields associated with the privacy label of the target data in the processing flow.

[0034] The above information is integrated to form the original data set of AI behavior trajectory fingerprint, and the final fingerprint value is generated through algorithms such as hashing or digital signatures. This fingerprint can uniquely identify the key elements of this AI processing. Then, the system logically associates this AI behavior trajectory fingerprint with the privacy label of the target data (for example, writing the fingerprint ID and label ID into the system database or metadata index at the same time) so that the data used in the processing and its permission attributes can be accurately traced back later.

[0035] Regarding S104 above: In specific implementation, at each stage or key step of AI processing, the AI ​​behavior trajectory fingerprint of the current execution process is compared with the privacy label of the target data in real time: if the trajectory fingerprint is consistent with the label and meets the usage restrictions, the current processing will continue; if it is found that the execution node or interface call recorded in the trajectory fingerprint is inconsistent with the allowable range specified by the label, the system will immediately block or switch to a stricter restriction and issue a warning. Dynamic access control ensures that the use of data is always consistent with its privacy requirements, thereby avoiding over-scope or unauthorized operations.

[0036] Regarding S105 above: In the specific implementation, after the AI ​​processing is completed, the final AI behavior trajectory fingerprint and the corresponding privacy label information are written into the audit module. The audit module can be deployed on the edge node or in the cloud to provide basic data support for subsequent compliance review, responsibility tracing or abnormality troubleshooting. The storage form can be a secure database or log file, and the timestamp and operator / device identity can be attached during storage to ensure the integrity and traceability of the information during subsequent verification.

[0037] For example, in a smart home environment, the security system is usually equipped with a variety of sensors, such as cameras, door and window sensors, and motion detectors. These devices continuously collect activity data of family members for real-time monitoring, intrusion detection, and abnormal behavior identification. In order to protect user privacy, each piece of collected video data is first pre-processed at the local camera device or the initial data collection gateway, and the corresponding privacy label is bound to the target data according to the preset sensitivity classification rules (for example, distinguishing whether it contains children or private conversations) and usage restriction policies (such as local monitoring only or monitoring only during a specific time period). The privacy label contains sensitivity information and usage restriction information, and is stored in a bundle with the video data in the form of encrypted metadata to ensure the consistency and security of the label and data during subsequent transmission and calculation processes.

[0038] Before the video data is prepared for AI processing, the pre-defined AI behavior space strategy database is queried according to the bound privacy label to determine the specific strategy and callable model or functional module for this AI processing. For example, for highly sensitive video data, intrusion detection processing can be allowed only on local devices or trusted edge nodes, and data cannot be uploaded to the cloud for analysis. At the same time, select an AI model version that has been optimized for privacy protection to ensure that the processing process meets privacy requirements.

[0039] During the execution of AI processing, key information involved in the processing is recorded, including the AI ​​model version used, the interface called, the identification of the execution node, the specific timestamp of the processing, and the key fields associated with the privacy label of the target data. This information is used to generate a unique AI behavior trajectory fingerprint through the SHA-256 hash algorithm to ensure the uniqueness and immutability of the fingerprint. The generated AI behavior trajectory fingerprint is logically bound to the privacy label of the target data through a database association table to ensure that the data used in each processing and its permission attributes can be accurately matched in the subsequent audit process.

[0040] At each key step of AI processing, the currently generated trajectory fingerprint is compared in real time with the usage restrictions defined in the privacy label. If the calling interface and execution node in the trajectory fingerprint are consistent with the usage restrictions of the privacy label, the AI ​​processing is allowed to continue; if it is found that the calling interface in the trajectory fingerprint attempts to access unauthorized functions, or the execution node is not within the allowed range, the processing flow will be interrupted immediately and switched to a stricter protection mode, such as only encrypting data output or reducing processing accuracy, and recording abnormal information for subsequent audits.

[0041] After AI processing is completed, the generated AI behavior trajectory fingerprint and the corresponding privacy label information are stored in the audit module deployed on the local smart home gateway device. The stored information includes trajectory fingerprint, privacy label information, whether the processing result is successfully executed, and whether there are any abnormalities.

[0042] For example, in the field of medical health, wearable devices such as smart bracelets and heart rate monitors continuously collect users' physiological data, including heart rate, blood pressure, and exercise volume, which are used for health monitoring and disease warning. However, these data are highly sensitive and involve personal privacy and health information, so strict privacy protection measures are required. First, in the privacy management module built into the wearable device, a privacy label is bound to each heart rate data according to the sensitivity of the data and the usage scenario. The privacy label includes sensitivity information, such as whether the heart rate data involves abnormal conditions (such as arrhythmia), and usage restriction information, such as personal health monitoring only or prohibited for third-party research. The privacy label is bound to the heart rate data in the form of encrypted metadata to ensure that the label and data are inseparable during data transmission and prevent the label from being tampered with or stripped.

[0043] Before the heart rate data is prepared for AI analysis, the system queries the AI ​​behavior space policy database based on the bound privacy label to determine the specific strategy and callable model or functional module for this AI processing. For example, for highly sensitive heart rate data, abnormal heart rhythm detection can be allowed only on the device or encrypted edge node, and data cannot be uploaded to the public cloud for analysis. At the same time, select an AI model version that meets privacy requirements, such as using a locally optimized lightweight model to reduce the need for data transmission and improve processing efficiency.

[0044] In this way, the target data can be managed in a combined way by combining "privacy labels" and "AI behavior trajectory fingerprints" in the IoT environment, achieving dynamic access control of the AI ​​processing flow, and leaving verifiable and traceable records in the audit module. The steps described in this implementation are all operational and feasible, and can be adjusted and expanded accordingly according to the actual IoT application scenarios and hardware conditions.

[0045] See also Figure 2 , Figure 2 A schematic diagram of the composition of an AI behavior trajectory provided for an embodiment of the present disclosure. In an optional implementation, the AI ​​behavior trajectory fingerprint can be generated by a hash algorithm or a digital signature algorithm when AI processing is performed on the target data. Specifically, when the system performs model reasoning, training or other computing operations on the target data, it first collects information related to this processing, including the model version used, the calling interface, the execution node, the timestamp, and the key fields associated with the privacy label of the target data. The model version is usually represented by a corresponding model identification number or version number to distinguish different algorithm models; the calling interface is used to record the service port or API name called during the processing; the execution node indicates the device or server identification where the AI ​​processing actually occurs; the timestamp accurately records the start or end time of this processing; the key fields associated with the privacy label of the target data are used to reflect information such as the data permissions or sensitivity levels used in this processing.

[0046] After all this information is integrated, a fingerprint value can be generated through a hash algorithm (such as SHA-256) or a digital signature algorithm (such as RSA or ECC-based signatures).

[0047] For example, if a hash algorithm is used, the above-mentioned key information can be combined into a character sequence or data structure, and then a hash value of a fixed length is calculated through a hash function, and it is saved locally or in a corresponding database as an AI behavior trajectory fingerprint.

[0048] For example, if a digital signature algorithm is used, the key information is first hashed, and then the hash value is signed using the system private key to obtain a signature result with anti-repudiation characteristics, further ensuring the immutability and source verifiability of the fingerprint. In the subsequent audit or traceability process, the hash value can be recalculated or the signature can be verified to check whether there is tampering or abnormal calls, thereby ensuring that each AI processing process can be traced back to its corresponding model version, call interface, execution node, and key information related to the target data privacy label, providing a reliable technical means for privacy protection in the IoT environment.

[0049] In an optional implementation, in order to further enhance the privacy protection of the target data, a dynamic access control link can be set during the AI ​​processing process, and the current AI behavior trajectory fingerprint and the privacy label of the target data can be compared in real time at each key step. When it is detected that the trajectory fingerprint is consistent with the sensitivity information and usage restriction information carried by the privacy label, it is determined that the current processing behavior is within the scope of authorization, thereby allowing the AI ​​processing process to continue; if the trajectory fingerprint does not match the privacy label, or if the current node or call interface is found in the authorization check to exceed the usage restrictions specified by the label, it means that the current processing behavior may be unauthorized or in violation of regulations, and the processing process can be interrupted immediately, or switched to a stricter protection mode, such as enabling a higher level of encryption strategy, outputting fuzzy results, or reducing data accuracy. At the same time, this abnormal information is recorded in the audit module for subsequent compliance review or risk investigation.

[0050] In this way, through this dynamic access control mechanism, it is possible to perform fine-grained control on data usage throughout the entire process of AI processing, achieve real-time monitoring and prevention of privacy risks, and ensure that each data processing is carried out within the legal and compliant scope.

[0051] See also Figure 3 , Figure 3 A schematic diagram of the contents of a privacy label provided in an embodiment of the present disclosure. As an optional implementation, the privacy label includes: The sensitivity level field is used to indicate the confidentiality level of the target data; The usage restriction field is used to describe the AI ​​processing scenarios in which the target data can be used; The compliance requirement field is used to describe the regulations or industry standards that the target data must comply with; The life cycle field is used to indicate the storage or usage period of the target data in the IoT environment.

[0052] In a specific implementation, in order to more accurately identify and manage the privacy attributes of target data in an IoT environment, multiple fields may be set in the privacy tag to comprehensively describe the security and compliance requirements of the data.

[0053] Among them, the sensitivity level field is used to indicate the confidentiality of the target data, which can be distinguished according to the sensitivity level of the data itself, such as confidential, restricted or public, so that corresponding protection strategies can be adopted according to different levels during subsequent AI processing.

[0054] The usage restriction field is used to describe the AI ​​processing scenarios in which the target data can be used, such as allowing reasoning to be performed only on local edge nodes or only by specific types of models, so that it can be accurately determined whether there is any unauthorized behavior during the dynamic access control stage.

[0055] The compliance requirement field is used to record laws, regulations, or industry standards related to the target data, such as compliance clauses related to medical, financial, or personal privacy protection. When data is migrated or processed, it can be used to verify whether the processing process meets the corresponding compliance requirements.

[0056] As an optional implementation, the compliance requirement field may include a regional ID (such as EU, US) or a regulatory ID (such as GDPR, HIPAA). When processing, the system will give priority to comparing the region to which the current node belongs (such as through IP address mapping, VPN export information, or edge node configuration). If a mismatch is found, further processing is prohibited or switching is required to a regional node that complies with the legal provisions.

[0057] For example, before executing AI processing, the system reads the regulatory ID in the privacy label. If it detects that GDPR needs to be met, but the GDPR compliance module is not enabled on the computing node or legal authorization is not obtained, the system will automatically refuse to process or switch to an edge node / cloud that meets GDPR requirements.

[0058] The life cycle field is used to indicate the storage or usage duration of the target data in the IoT environment. This field can be used to set a data expiration policy or an automated destruction mechanism. Once the specified time limit is exceeded, the system will promptly trigger the cleanup process and update or revoke the privacy label of the data accordingly, thereby technically ensuring that the data is not stored or used indefinitely, further reducing the risk of privacy leakage.

[0059] For example, in order to more accurately identify and manage the privacy attributes of target data in the IoT environment, after the target data from the terminal device is collected, a privacy label containing four main fields will be automatically bound to each piece of data. First, the sensitivity level field indicates the confidentiality of the target data. The system usually determines whether the data is confidential, restricted or public based on established classification rules, and records the result in the label. For example, for some medical data that may contain personal biometrics, the system will mark it as confidential so that it can only be processed on nodes with a higher security level in the subsequent computing migration process.

[0060] Secondly, the usage restriction field is used to describe the AI ​​processing scenarios in which the target data can be used. In specific implementations, the filling of this field can be refined according to different application modules. For example, for industrial monitoring data, real-time fault prediction is allowed on edge nodes, but the data is prohibited from being used for commercial competition analysis; or for home surveillance video data, only intrusion detection is allowed on local devices to avoid uploading to the public cloud for face recognition. In this way, when the AI ​​processing module receives the data, it can determine whether it meets the preset usage scope by comparing this field, thereby triggering subsequent access control logic.

[0061] Furthermore, the compliance requirement field is used to describe the laws, regulations and industry standards that the target data must comply with in a specific industry or region. Taking the medical scenario as an example, if the target data involves personal health indicators, the system will record the relevant medical privacy protection regulations in this field to ensure that the processing flow can verify whether it complies with local medical data compliance requirements. Once it is found during the review that the data processing steps do not meet the laws and regulations corresponding to this field, the system will trigger an abnormal warning and immediately interrupt or adjust the current AI processing process.

[0062] In addition, the lifecycle field is used to clarify the storage or usage duration of the target data in the IoT environment. When the label is generated, the availability period is set according to the data type and security policy. For example, some video data can only be kept for two weeks for temporary backtracking and anomaly analysis. After expiration, this part of the data must be automatically deleted or desensitized, and the privacy label must be synchronously updated to an unavailable state. In this way, when subsequent AI processing tasks try to access expired data, they will find that the data lifecycle has ended through dynamic authorization checks, thereby refusing to call or re-authorizing the data.

[0063] In this way, through the coordination of the above four core fields, the sensitivity, usage scenarios, compliance requirements and storage duration of each piece of data can be finely controlled in the IoT environment, reducing the risk of data abuse and leakage.

[0064] In an optional implementation, in order to timely monitor and prevent possible privacy violations in various AI processing links in the IoT environment, the present disclosure divides the AI ​​behavior space into a data preprocessing stage, a model reasoning stage, a result output stage, and a model update stage in advance, and dynamically checks the privacy label and AI behavior trajectory fingerprint in each stage, so as to timely block the continued use of the target data when the privacy label constraints are not met.

[0065] In specific implementation, when an IoT terminal device or edge node receives a batch of target data that will be involved in AI analysis, the system will first enter the data preprocessing stage to clean, convert the format, or simply aggregate the target data. Before starting these operations, the system will read the bound privacy label, verify whether the sensitivity level field and usage restriction field allow the preprocessing task to be performed on the node, and compare the execution node, timestamp and other information recorded in the currently generated AI behavior trajectory fingerprint to confirm that this preprocessing request is legal and complies with the established policy. If both the label and fingerprint checks pass, the system will continue data preprocessing, otherwise subsequent operations will be immediately blocked.

[0066] During the model inference phase, the system uses the pre-processed target data as input and calls the AI ​​model deployed in advance on the edge node or cloud for inference. In order to ensure that there is no unauthorized use or illegal transmission of sensitive data at this stage, the system will re-evaluate the privacy label before formally executing the inference. If the label's usage restriction field indicates that inference can only be performed in a specific environment or on a specific model, the system needs to detect whether the model version, execution node and other information called this time are consistent with the label. At the same time, the system will check the current AI behavior trajectory fingerprint to see whether the call interface and timestamp recorded therein are consistent with the expected information in the model inference phase. If there is a mismatch at any point, the system will immediately interrupt the inference process and write the abnormal operation into the audit module. Otherwise, the model inference will be executed normally until the inference result is obtained.

[0067] During the result output phase, the system will return the result data generated by the model reasoning to the upper-level application or store it in the database. In order to prevent the inference results containing sensitive information from being leaked or disseminated without authorization, the system will verify again based on the privacy label and AI behavior trajectory fingerprint before output. If the privacy label indicates that the result needs to be desensitized or encrypted before it can be output, the system will call the corresponding security module to process the result, and then add the timestamp and processing method of this operation to the AI ​​behavior trajectory fingerprint to ensure that the compliance verification of privacy requirements has been completed before the result is output; if it is found that the result to be output contains sensitive content that exceeds the scope allowed by the label, it will trigger an emergency restriction mechanism, such as only outputting summary information or interrupting the result transmission, and recording the violation attempt to the audit system for subsequent tracing of responsibility or compliance review.

[0068] During the model update phase, the system will use the newly acquired data features or annotation information to update the AI ​​model itself to improve the accuracy or applicability of the model in subsequent inference tasks. Since the model update process often involves secondary use of the original data and modification of model parameters, the system will still review it according to the scenario restrictions recorded in the privacy label and AI behavior trajectory fingerprint. If the privacy label restricts the target data from being used for training or self-learning, the system will immediately terminate the operation when a model update request is detected to prevent the model from using restricted data; at the same time, by comparing with the trajectory fingerprint, it can be further confirmed whether there is an out-of-range node attempting to perform an update operation. Once discovered, the system will immediately cut off the update process and record the abnormal event.

[0069] In this way, at each key stage of the entire AI behavior space, the system can strictly control privacy labels and AI behavior trajectory fingerprints for fine-grained control, thereby effectively avoiding potential risks of sensitive data due to erroneous processes, unauthorized access or illegal use.

[0070] In an optional implementation, in order to effectively protect the target data and the privacy labels and AI behavior trajectory fingerprints it carries during computational migration, the system will first encrypt the target data, its privacy labels, and the currently generated AI behavior trajectory fingerprints before migrating the target data from the IoT terminal device to the edge node or cloud platform.

[0071] In the specific implementation, the terminal device will call the pre-configured security module to perform symmetric or asymmetric encryption operations on the data content and metadata information to be transmitted, and can also digitally sign the structured information containing privacy labels and trajectory fingerprints to ensure that it is not tampered with or forged during transmission. After encryption is completed, the terminal device sends these encrypted data packets to the designated edge node or cloud platform through a secure communication protocol (such as TLS or a custom encrypted tunnel).

[0072] When the edge node or cloud platform receives these data packets, it will first extract and decrypt the attached privacy label and AI behavior trajectory fingerprint, and use the digital signature to verify its integrity and authenticity. Based on the results of decryption and verification, the system determines whether the security and compliance requirements of this AI processing task are met. If after decryption, it is found that the privacy label explicitly restricts its processing at the current node, or the verification process detects abnormal conditions such as fingerprint and label mismatch, signature verification failure, etc., the system will refuse to continue AI processing, and will determine this transmission as a security anomaly, triggering the audit module to record relevant information; if everything is verified normally and the usage restriction field in the label allows inference or training at the receiving end, the system will load the target data and its privacy label and fingerprint information into the AI ​​processing environment, and continue to perform the next inference, result output or model update operations according to the established process.

[0073] In this way, through encryption and decryption verification during the migration process, the risk of sensitive data leakage caused by network transmission or cross-node collaboration can be minimized, and at the same time, strict review of the compliance use of data at the receiving end is guaranteed, thereby effectively enhancing the security and privacy protection level of computing migration in the IoT environment.

[0074] As an optional implementation, the content stored in the audit module includes: The AI ​​behavior trajectory fingerprint corresponding to each AI processing process; Privacy label information associated with the AI ​​behavior trajectory fingerprint; The execution results of dynamic authorization or access control; Records of detected anomalies or violations; The audit module is deployed on edge nodes or cloud platforms to support verification and violation forensic analysis.

[0075] In an optional implementation, in order to systematically record and trace the key operations of the target data during the AI ​​processing process, the present disclosure deploys a dedicated audit module on the edge node or cloud platform for centralized storage and management of key audit information.

[0076] In the specific implementation, after completing an AI processing task, a series of information related to the task will be written into the secure storage area of ​​the audit module. This includes the AI ​​behavior trajectory fingerprint corresponding to each AI processing process, which is generated by the model version, call interface, execution node, timestamp involved in this processing, and key fields associated with the privacy label of the target data through a hash algorithm or a digital signature algorithm.

[0077] Furthermore, after receiving the fingerprint, the audit module will logically associate it with the corresponding privacy label information to ensure that the data used in this processing, the data sensitivity and the permitted scope of use can be clearly reflected in subsequent retrieval.

[0078] In addition, in order to achieve a comprehensive review of process compliance, the execution results of this dynamic authorization or access control will also be written into the audit module.

[0079] In specific implementation, if the risk of unauthorized calls or out-of-scope usage is detected at a certain stage of AI processing, the system will simultaneously upload the exception or violation records to the audit module, retaining detailed information such as the time point, execution node ID, and the specific interface or behavior that triggered the violation.

[0080] In this way, when administrators or compliance reviewers need to conduct verification or forensic analysis of violations in the future, they can query the entire track of this AI processing through the audit module, including fingerprint information, privacy labels, authorization verification status, and any abnormal warnings.

[0081] For example, the audit module has high security and access control capabilities. Only managers with corresponding permissions can read and write to it, thus avoiding the risk of data leakage or malicious tampering of audit logs. By managing these elements in a unified manner in the audit module, it can provide a strong basis for post-tracing and responsibility division for the privacy protection mechanism in the IoT environment, taking into account the needs of real-time protection and post-compliance audit.

[0082] In an optional implementation, in order to enhance the credibility and tamper resistance of the audit data, the association information of the AI ​​behavior trajectory fingerprint and the privacy label can be stored in a blockchain or a consortium chain side chain.

[0083] In specific implementation, after the edge node or cloud platform completes the audit record of the AI ​​processing process, the behavior trajectory fingerprint, associated privacy label information, key timestamps and other metadata generated by this processing will be packaged into transaction or block candidate information, and distributed verification will be carried out in the alliance chain side chain or blockchain network through the pre-deployed consensus mechanism. Each node participating in the consensus verifies the validity and consistency of the transaction or block candidate information according to the preset consensus algorithm (such as PBFT, Raft or PoW, etc.). Once a sufficient number of nodes confirm, the information is written to the chain and an independent block hash is generated, forming an unalterable record of the audit data.

[0084] Furthermore, in the subsequent compliance review or data tracing process, any node can verify on the blockchain or consortium chain side chain whether the corresponding behavior trajectory fingerprint is consistent with the information originally stored, and confirm whether it is consistent with the privacy label associated information on the blockchain or consortium chain side chain.

[0085] For example, if there is an attempt to tamper with or delete on-chain records, other nodes will detect data inconsistency during the consensus process and refuse to accept the modification.

[0086] In this way, the present disclosure not only retains local audit logs on edge nodes or cloud platforms, but also uses distributed ledgers to further enhance the authenticity and security of audit data, effectively improving the transparency and credibility of AI behavior auditing and privacy protection in the IoT environment.

[0087] As an optional implementation, when generating the AI ​​behavior trajectory fingerprint, it also includes: Generating sub-fingerprints for each of the multiple sub-steps in the AI ​​processing process, and merging the sub-fingerprints into the final AI behavior trajectory fingerprint after the processing is completed; The sub-fingerprint is used to distinguish the processing information of different sub-steps during the audit phase.

[0088] In an optional implementation, in order to record the AI ​​processing process in more detail and distinguish each link in the audit stage, when generating the AI ​​behavior trajectory fingerprint, sub-fingerprints are generated for multiple sub-steps in the AI ​​processing process, and then these sub-fingerprints are merged into the final AI behavior trajectory fingerprint after the overall processing is completed.

[0089] In the specific implementation, the AI ​​processing flow is first divided into several logical sub-steps, such as data cleaning preprocessing, feature extraction, model inference, and result verification, and before each sub-step starts, information such as the currently called model version, execution node, usage interface, timestamp, and key fields associated with the privacy label are collected.

[0090] For each sub-step, the system generates a sub-fingerprint based on this information and caches it locally or records it in an intermediate database. After the entire AI processing flow is completed, the system merges each sub-fingerprint in chronological order or execution order, and forms a complete AI behavior trajectory fingerprint through hash aggregation or signature aggregation algorithm, which is finally stored in the audit module or recorded on the chain.

[0091] In addition, to prevent attackers from bypassing the fingerprint generation module, a secure firmware / trusted execution environment (TEE) can be built into the edge node or terminal device. Only through this environment can the AI ​​processing interface be called; when entering each sub-step, the system will obtain the real input data and label information from the TEE, perform hash or digital signature operations on it, and generate / update the sub-fingerprint. If any link fails to generate a log or fingerprint, it is considered a processing failure and an alarm is triggered.

[0092] In this way, in the subsequent audit stage, the administrator can trace back each sub-fingerprint to accurately locate the resources and data used in each step of the processing, as well as whether there are any illegal calls or unauthorized access.

[0093] For example, if signs of abnormality are found in an intermediate sub-step, the system can immediately trigger an alarm and mark the source of the abnormality when merging fingerprints, making the audit and traceability of the entire AI processing chain more refined and efficient.

[0094] In an optional implementation, the IoT environment is composed of multiple levels of devices and nodes, and performs privacy protection and AI processing of target data in a distributed collaborative manner.

[0095] In the specific implementation, first, multiple resource-constrained terminal devices collect environmental data or user data locally, and bind privacy labels after data generation or preprocessing to ensure that sensitive information and usage restriction information can be strictly followed in subsequent links. Due to limited computing power and storage capacity, these terminal devices usually only undertake primary or partial AI processing functions, such as simple filtering, data denoising, and local security encryption. When encountering complex tasks or high load requirements, they will migrate data or part of the computing tasks to more powerful edge nodes for execution.

[0096] At the edge node level, the system will execute at least part of the calculation process of the stage and generate the AI ​​behavior trajectory fingerprint according to the configuration of the AI ​​behavior space. Specifically, after receiving the target data from the terminal device, the edge node will first verify the sensitivity and usage restrictions of the data based on the privacy label, determine the callable model version and processing interface, and then perform operations such as model reasoning, result output or model update. Whenever a certain stage is completed, the edge node will collect information such as the current execution node, call interface, timestamp, etc., and then combine the key fields associated with the privacy label to generate the AI ​​behavior trajectory fingerprint through a hash or digital signature algorithm. If the edge node detects signs of unauthorized access or out-of-scope use, it will immediately terminate the processing and submit the abnormal information to the audit module or cloud for recording.

[0097] In the cloud platform, the system is responsible for the unified management of trajectory fingerprints and privacy labels generated on terminal devices and edge nodes, and performs centralized or distributed audit operations. The cloud platform builds an overall data usage chain by receiving trajectory fingerprints, authorization verification results and exception logs from each node, so as to conduct compliance review or traceability analysis for any processing process when necessary. When the cloud decides to further participate in AI processing or conduct higher-level modeling and training, it will also restrict whether data upload is allowed and the AI ​​models that can be called according to the usage restriction fields of the privacy label and the system's dynamic access control policy. At the same time, the cloud platform will record the received fingerprint and label mappings in the security audit database or distributed ledger, and is responsible for authority management. Only authorized operation and maintenance personnel or auditors can view or analyze the relevant logs.

[0098] In addition, in actual deployment, each edge node or terminal device uses NTP (Network Time Protocol) or GPS reference clock for periodic synchronization, or uses the standard timestamp of block records when storing blocks on the blockchain side chain to ensure consistency in the order of events within a tolerable range.

[0099] For example, when verifying fingerprints and tags, a preset time offset (for example, ±5 seconds) is allowed for the comparison of timestamps. If the processing sequence can still be confirmed within this deviation, the system considers it valid. Otherwise, a secondary verification or alarm is triggered.

[0100] In this way, the terminal devices, edge nodes and cloud platforms can work together within their respective computing power ranges, which can not only meet the needs of real-time processing and computing migration in the IoT scenario, but also perform systematic privacy protection according to the methods described above. Each node in this environment has a clear role positioning and data processing permissions, ensuring that the entire process from data collection to final audit is smoothly carried out under the joint constraints of privacy labels and AI behavior trajectory fingerprints, thereby comprehensively improving the security and reliability of IoT applications.

[0101] Those skilled in the art can understand that in the above method of the specific implementation, the order of writing each step does not mean a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic. It should be understood that determining B based on A does not mean determining B based only on A, but B can also be determined based on A and / or other information.

[0102] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0103] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

Claims

1. A privacy protection method based on computational migration and AI behavior trajectory analysis, characterized in that: For use in an Internet of Things environment, the method comprises: Binding a privacy label to target data in the Internet of Things; wherein the privacy label includes: sensitivity information and usage restriction information of the target data; Before performing AI processing on the target data, determining a processing strategy in an AI behavior space based on the privacy label; wherein the AI ​​behavior space includes: a data preprocessing stage, a model reasoning stage, a result output stage, and a model update stage; During the execution of the AI ​​processing, based on the model version, the calling interface, the execution node, the timestamp, and the information associated with the privacy label of the target data, an AI behavior trajectory fingerprint is generated, and the AI ​​behavior trajectory fingerprint is associated with the privacy label; According to the AI ​​behavior trajectory fingerprint and the privacy label, dynamic access control is performed on the AI ​​processing process; After the AI ​​processing is completed, the generated AI behavior trajectory fingerprint and the corresponding privacy label information are stored in the audit module for verification and tracing.

2. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 1 is characterized in that: The AI ​​behavior trajectory fingerprint is generated by a hash algorithm or a digital signature algorithm; The AI ​​behavior trajectory fingerprint includes: model version, calling interface, execution node, timestamp, and key fields associated with the privacy label of the target data.

3. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 2 is characterized in that: The dynamic access control includes: In response to the AI ​​behavior trajectory fingerprint matching the privacy label, allowing the AI ​​processing process to continue; In response to the AI ​​behavior trajectory fingerprint not matching the privacy label or detecting out-of-range use, the AI ​​processing process is interrupted or the protection mode is switched, and abnormal information is recorded.

4. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 3 is characterized in that: The privacy labels include: The sensitivity level field is used to indicate the confidentiality level of the target data; Usage restriction field, used to describe the AI ​​processing scenario in which the target data is used; The compliance requirement field is used to describe the regulations or industry standards that the target data complies with; The life cycle field is used to indicate the storage or usage period of the target data in the IoT environment.

5. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 4 is characterized in that: Also includes: In each stage included in the AI ​​behavior space, the privacy label and the AI ​​behavior trajectory fingerprint are respectively given a dynamic check to restrict the continued use of the target data when the privacy label constraint is not met.

6. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 5 is characterized in that: When generating the AI ​​behavior trajectory fingerprint, it also includes: Generating sub-fingerprints for each of the multiple sub-steps in the AI ​​processing process, and merging the sub-fingerprints into the final AI behavior trajectory fingerprint after the processing is completed; The sub-fingerprint is used to distinguish the processing information of different sub-steps during the audit phase.

7. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 6, characterized in that: Also includes: When migrating the target data from the IoT terminal device to the edge node or cloud platform, the privacy label and the AI ​​behavior trajectory fingerprint of the target data are encrypted and transmitted, and after the encrypted information is decrypted and verified at the receiving end, a decision is made as to whether to continue the AI ​​processing process.

8. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 7, characterized in that: The contents stored in the audit module include: The AI ​​behavior trajectory fingerprint corresponding to each AI processing process; Privacy label information associated with the AI ​​behavior trajectory fingerprint; The results of dynamic authorization or access control execution; detected anomalies or violation records; The audit module is deployed on edge nodes or cloud platforms to support verification and violation forensic analysis.

9. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 8, characterized in that: Also includes: The association information of the AI ​​behavior trajectory fingerprint and the privacy label is stored in the blockchain or the consortium chain side chain, and the association information is verified and recorded through a consensus mechanism.

10. The privacy protection method based on computational migration and AI behavior trajectory analysis according to claim 9 is characterized in that: The IoT environment includes: Multiple resource-constrained terminal devices use the method to bind privacy labels to the collected target data; At least one edge node executes at least a portion of the stages in the AI ​​behavior space and generates the AI ​​behavior trajectory fingerprint.

Images