A method and system for adaptive privacy protection of urban data based on dynamic security classification

Abstract

This invention relates to the field of data security protection technology, and in particular to a method and system for adaptive privacy protection of urban data based on dynamic security classification. The method involves multi-source data semantic perception and dynamic classification based on preprocessed data, including semantic feature extraction of data content based on knowledge graphs, dynamic risk entropy calculation based on scene and environment perception, and automatic mapping and label generation of data security levels. Based on the dynamic classification results, source-end adaptive privacy perturbation is applied to the data, including policy mapping and lightweight desensitization based on security levels, adaptive allocation of privacy budget based on spatiotemporal density perception, and random noise injection and data release. By constructing a multi-source data semantic perception and dynamic risk entropy model, the method overcomes the rigidity of traditional rule-based classification, achieving accurate classification and dynamic risk management of massive amounts of heterogeneous urban data.

Description

Technical Field

[0001] This invention relates to the field of data security protection technology, and in particular to an adaptive privacy protection method and system for urban data based on dynamic security classification. Background Technology

[0002] Existing urban data protection systems still face severe challenges in practice: First, protection strategies are rigid and lack dynamic perception capabilities. Traditional security measures often employ a "one-size-fits-all" static classification and encryption strategy, failing to perceive changes in data risk under different application scenarios and spatiotemporal environments. This leads to overprotection of low-risk data, impacting efficiency, or insufficient protection of high-risk data in specific scenarios. Second, privacy and usability are difficult to balance. When faced with massive amounts of heterogeneous data, conventional anonymization or encryption techniques can easily damage the statistical characteristics of the data, resulting in data "preserving privacy but losing value," making it difficult to support accurate urban operation analysis. Third, trust gaps exist in cross-domain collaboration. In joint modeling and collaborative computing, there is a lack of effective implementation mechanisms for "data remaining stationary while the model moves." Frequent transfers of plaintext data make it difficult to separate data ownership from usage rights, easily leading to secondary leakage and misuse of sensitive information.

[0003] To address the aforementioned issues, domestic and international research has primarily focused on enhanced access control, differential privacy algorithms, and privacy-preserving computation frameworks. Traditional role-based access control (RBAC) struggles to handle the complex, fine-grained permission requirements of multi-source data; simple differential privacy mechanisms often face usability collapse due to noise accumulation when processing high-dimensional urban data; while homomorphic encryption and multi-party secure computation offer high security, their computational overhead is enormous, making them difficult to directly apply to high-concurrency real-time urban decision-making scenarios. Consequently, existing technologies generally suffer from coarse-grained protection, significant computational performance bottlenecks, and a lack of end-to-end linkage mechanisms.

[0004] The current urban data security protection system mostly relies on rule-based static classification, fixed-strength desensitization algorithms or traditional boundary defense mechanisms, which have the following prominent defects: (1) Security classification is separated from application scenarios and lacks dynamic perception capabilities. Existing solutions mostly use predefined static labels to classify data in one go, which cannot perceive the risk evolution of data under different business scenarios, spatiotemporal environments and fusion associations, resulting in the contradiction of "insufficient protection of low-risk data in sensitive scenarios" or "overprotection of high-risk data in ordinary scenarios", which is difficult to adapt to the complex and ever-changing urban governance needs; (2) The protection mechanism is rigid and lacks adaptive adjustment capabilities. Traditional technologies often use a "one-size-fits-all" uniform desensitization rule or fixed-parameter encryption algorithm, which cannot dynamically adjust the differential privacy budget, encryption granularity or access control strategy according to the real-time security level of the data, resulting in serious distortion of data statistical characteristics or extreme waste of computing resources, making it difficult to achieve the optimal balance between privacy protection strength and data availability; (3) The risk of privacy leakage in cross-domain collaborative computing is high and the bottleneck of computing efficiency is prominent. When it comes to joint modeling and real-time querying involving multiple departments, there is a lack of efficient implementation methods to ensure that "data is available but not visible." Plaintext transmission carries the risk of being stolen or misused. Full homomorphic encryption or general federated learning technologies have huge computational overhead and cannot meet the stringent real-time requirements of scenarios such as urban emergency response.

[0005] Therefore, there is an urgent need to propose an adaptive privacy protection method and system for urban data based on dynamic security classification. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method and system for adaptive privacy protection of urban data based on dynamic security classification.

[0007] In a first aspect, the present invention provides an adaptive privacy protection method for urban data based on dynamic security classification, which adopts the following technical solution:

[0008] An adaptive privacy protection method for urban data based on dynamic security classification includes:

[0009] Acquire multi-source data;

[0010] Data preprocessing is performed based on the acquired multi-source data;

[0011] Multi-source data semantic perception and dynamic classification are performed based on preprocessed data, including semantic feature extraction of data content based on knowledge graph, dynamic risk entropy calculation based on scene and environment perception, and automatic mapping and tag generation of data security level.

[0012] Based on the dynamic rating results, the data undergoes source-end adaptive privacy perturbation, including security level-based policy mapping and lightweight desensitization, spatiotemporal density-aware privacy budget adaptive allocation, and random noise injection and data publication.

[0013] The perturbation-processed data is encrypted and stored based on a hierarchical strategy, including the automatic generation of access policies based on hierarchical mapping, encryption processing based on improved CP-ABE, and distributed erasure coding storage and integrity verification.

[0014] The stored data is protected by multiple layers of security and privacy, including privacy set intersection based on RSA blind signature, adaptive switching of federated protocols based on security level, and secure aggregation and global model update.

[0015] Selective closed-state queries and responses are performed based on the protected data.

[0016] Secondly, a city data adaptive privacy protection system based on dynamic security classification includes:

[0017] The data acquisition module is configured to acquire data from multiple sources.

[0018] The preprocessing module is configured to perform data preprocessing based on the acquired multi-source data;

[0019] The dynamic classification module is configured to perform multi-source data semantic perception and dynamic classification based on preprocessed data, including data content semantic feature extraction based on knowledge graph, dynamic risk entropy calculation based on scene and environment perception, and automatic mapping and tag generation of data security level.

[0020] The perturbation module is configured to perform source-end adaptive privacy perturbation on the data based on the dynamic classification results, including security level-based policy mapping and lightweight desensitization, spatiotemporal density-aware privacy budget adaptive allocation, and random noise injection and data publishing.

[0021] The storage module is configured to encrypt and store the perturbation-processed data based on a hierarchical strategy, including automatic generation of access strategies based on hierarchical mapping, encryption processing based on improved CP-ABE, and distributed erasure coding storage and integrity verification.

[0022] The protection module is configured to provide multi-layered security and privacy protection for the stored data, including privacy set intersection based on RSA blind signature, adaptive switching of federated protocols based on security level, and secure aggregation and global model update.

[0023] The response module is configured to perform selective encrypted queries and responses based on the protected data.

[0024] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned adaptive privacy protection method for urban data based on dynamic security classification.

[0025] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the aforementioned adaptive privacy protection method for urban data based on dynamic security classification.

[0026] In summary, the present invention has the following beneficial technical effects:

[0027] Compared with existing technologies, the urban data adaptive privacy protection method and system based on dynamic security classification proposed in this invention have the following significant advantages: It breaks through the limitations of static classification by constructing a multi-source data semantic perception and dynamic risk entropy model, overcoming the rigidity of traditional rule-based classification, and achieving accurate classification and dynamic risk control of massive urban heterogeneous data; it resolves the contradiction between privacy and utility by utilizing an adaptive perturbation strategy based on source-end spatiotemporal density perception, dynamically adjusting noise according to environmental density, significantly reducing the risk of sensitive information reconstruction while maximizing the effective utilization value of data; it achieves "usable but invisible" cross-domain collaboration by integrating attribute encryption and adaptive federated learning technology to construct a secure collaborative ecosystem where "data remains stationary while the model moves," effectively breaking down data barriers between departments; and it balances computational efficiency and reliable traceability by significantly reducing system query response latency through a level-triggered selective encrypted query and blockchain evidence storage mechanism, while ensuring the immutability and auditability of the entire process, significantly improving the intelligence and security level of urban data governance. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of an adaptive privacy protection method for urban data based on dynamic security classification driven according to Embodiment 1 of the present invention;

[0029] Figure 2 This is a schematic diagram comparing different methods in Embodiment 1 of the present invention;

[0030] Figure 3 This is a comparison chart of data utility under different privacy protection levels in Embodiment 1 of the present invention. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings.

[0032] Example 1

[0033] Reference Figure 1 This embodiment of an adaptive privacy protection method for urban data based on dynamic security rating includes:

[0034] To address key issues in the entire lifecycle of urban data, such as static security classification, rigid protection strategies, and high privacy leakage risks in cross-domain collaborative computing, this invention proposes an adaptive privacy protection method and system for urban data based on dynamic security classification. The system's overall design consists of six core modules: a multi-source data acquisition and access module, used to achieve standardized access and unified cataloging of heterogeneous data from multiple fields such as government affairs, transportation, environment, emergency response, and the Internet of Things, completing data format recognition, interface parsing, and metadata registration; a multi-source data semantic perception and dynamic classification module, which uses domain knowledge graphs and NLP technology to extract data content features, establishes a joint risk model based on real-time application scenarios and environmental characteristics, calculates dynamic risk entropy values, and automatically assigns corresponding data security levels; a source-end adaptive privacy perturbation module, which automatically matches differentiated preprocessing strategies according to security levels, performs k-anonymization desensitization on low-level data, and adaptively adjusts the differential privacy budget and noise injection amplitude based on spatiotemporal density characteristics for high-level data, balancing data availability and privacy at the source; and a hierarchical strategy generation module. Together with the encrypted storage module, an automatic mapping mechanism is constructed from security level to access control logic. It automatically generates encryption policies based on Ciphertext Policy Attribute Base (CP-ABE) to achieve fine-grained access control and ciphertext persistence at the storage layer. The cross-domain joint protection and protocol switching module is designed for multi-department collaborative analysis scenarios. Based on the highest security level of the data involved in the computation, it adaptively switches between lightweight federated learning and a vertical federated mechanism based on secret sharing, completing cross-domain model training and parameter aggregation without exchanging original data. The selective encrypted query and response module is designed for real-time retrieval needs. It triggers the corresponding encrypted computing engine based on the security level of the query object, supports ciphertext retrieval and computation in a fully homomorphic encrypted environment, and is combined with a full-process blockchain audit mechanism to ensure privacy protection, immutability, and full traceability of the data query process.

[0035] S1. Multi-source data acquisition and access module

[0036] In the construction of the City Brain data platform, the raw data comes from government business systems, IoT sensing devices, video surveillance networks, transportation platforms, and external open data sources such as meteorology, environment, and energy. These data types include structured tabular data, semi-structured log data, unstructured text, and video streams, characterized by heterogeneous formats, varying spatiotemporal granularity, and large differences in update frequency. Without standardized processing and unified modeling, these data will severely impact the performance and reliability of subsequent data fusion, index optimization, and intelligent analysis. Therefore, this module aims to achieve unified access, format conversion, spatiotemporal registration, and structured representation of multi-source heterogeneous data, forming a standardized and computable data input system. The overall processing flow includes the following three parts:

[0037] 1) Access and standardized processing of multi-source heterogeneous data: A unified data access framework is adopted for data streams from different sources. An interface adaptation layer is used to achieve multi-protocol parsing (HTTP, MQTT, Kafka, FTP, etc.) and data format recognition. Let the original input dataset be... D for:

[0038] ,

[0039] in, d i Indicates the first i One original data record, s i This represents the data source identifier, used to distinguish different source systems or sensing devices. f i Indicates the first i The original set of fields contained in each data record. t i For timestamps, a unified data standardization mapping function is defined for both structured and semi-structured data to obtain standardized feature values:

[0040] ,

[0041] in, Indicates the first i The first data record j The original values ​​of each field, and These represent the mean and standard deviation of the field in the global data, respectively. This normalization process maps the numerical distributions of different data sources to a standard normal space, eliminating dimensional differences and scale shifts between different data sources, and ensuring the comparability and convergence stability of input features in subsequent modeling stages.

[0042] For unstructured text (such as government documents and public opinion reports), key fields are extracted and entities are standardized using Natural Language Processing (NLP) techniques. Semantic tags are then extracted and uniformly named to achieve semantic-level fusion. For video streams or image data, frame extraction and time indexing are used to transform unstructured information into structured descriptions. The final result is a data object in a unified format.

[0043] ,

[0044] in, ID i Indicates the data object number. Type i Represents data types (such as tables, video frames, logs, images, etc.). Feature i This represents a vectorized feature description generated by an encoding network or feature extraction model. Time i Indicates the time index of the data. Location i This indicates the spatial location information of the data (such as latitude and longitude coordinates).

[0045] 2) Spatiotemporal alignment and data synchronization: Due to the different acquisition cycles of various sensing devices and business systems, data often suffers from time asynchrony and spatial offset. To achieve spatiotemporal consistency across data sources, a synchronization algorithm based on time windows and a spatial coordinate transformation mechanism are introduced. Time synchronization employs a linear interpolation method, based on the target time point. t * Estimate the data from two adjacent time points:

[0046] ,

[0047] in, x t and x t+1 They are time points respectively t and t The data value at +1, t * This indicates the target synchronization time. This ensures that multi-source data can be aligned under a unified time reference. For spatial alignment, for spatial data from different coordinate systems (such as Gaode coordinates GCJ-02, WGS-84, and local city coordinates), a two-dimensional affine transformation model is used to achieve coordinate unification.

[0048] ,

[0049] Where (x,y) and (x',y') are the geographical locations in the original and target coordinate systems, respectively, and the matrix parameter a 11 a12 a 21 a 22 t represents the rotation and scaling parameters. x , t y This represents a translation vector used to adjust spatial offset. This transformation enables spatial unification of data from multiple urban departments.

[0050] 3) Data integrity verification and missing value completion: During the long-term collection of government and sensor data, missing records or discontinuous sampling often occur due to factors such as network latency, device offline, and interface anomalies. To ensure the training and real-time calculation effects of subsequent models, a spatiotemporal weighted interpolation method is used for data completion. Let's assume a certain monitoring point... i In time t Missing values ​​exist Its neighborhood effective data set is Then the missing values ​​are estimated as follows:

[0051] ,

[0052] in, Indicates the monitoring point i In time t The estimated value, Representing neighborhood points j Valid data values ​​at the same time, p i and p j Let i and j represent the spatial locations of monitoring points i and j, respectively. p i - p j || represents the Euclidean distance. To prevent the use of tiny constants with a denominator of zero. Weights. w ij The weighted smooth interpolation of neighboring points decreases as distance increases, ensuring the continuity and consistency of data spatially. This leverages the spatiotemporal similarity of the neighborhood for smooth completion, guaranteeing data continuity and structural integrity. Simultaneously, to achieve real-time integrity monitoring of the data stream, an integrity detection function is introduced.

[0053] ,

[0054] in, C ( t ) indicates time t Data integrity metrics N valid ( t N represents the number of data points successfully collected at that moment. expected ( t) represents the theoretical number of data points that should be collected, and δ represents the threshold. C ( t When the threshold is less than δ (typically 0.90), the system automatically triggers a data re-acquisition or asynchronous repair process to ensure the stability and reliability of the input data to the middle platform.

[0055] S2. Multi-source data semantic awareness and dynamic classification module

[0056] After completing the access and standardization of multi-source heterogeneous data, the primary challenge facing the City Brain data platform is how to accurately identify the sensitivity of the data and dynamically classify it. This module aims to leverage domain knowledge graphs and natural language processing technologies to deeply analyze the semantic features of data content, construct a joint risk model based on real-time environmental context, calculate the dynamic risk entropy value of the data, and automatically assign corresponding security levels (L1-L4), providing a basis for subsequent privacy perturbation and encryption strategies. The module consists of three parts: semantic feature extraction, dynamic risk entropy calculation, and security level mapping and label generation.

[0057] 1) Knowledge graph-based semantic feature extraction of data content, targeting standardized data objects. O i First, we will construct a city data security knowledge graph covering areas such as government affairs, transportation, and policing. G KG =( E , R ),in E For a collection of entities, R This involves defining relationships between entities. Using a BERT-based Named Entity Recognition (NER) and relation extraction model, data objects are... O i The feature vectors in the dataset are mapped to the knowledge graph, and their intrinsic sensitivity is calculated. Data objects are defined. O i Content Sensitivity Index S content ( i ):

[0058] ,

[0059] in, e k Indicates the first extracted from the data k There are entities, I(·) is an indicator function (1 if the entity exists, 0 otherwise), W(e k ) represents the pre-defined basic sensitivity weights for this entity in the knowledge graph. asso ( e k ) represents an entity e k The degree of correlation in the graph, This is the correlation enhancement coefficient. The correlation term amplifies the combined privacy risks (e.g., the sensitivity of the combination of "name + trajectory" is higher than the sum of the two individual terms), thereby achieving a deeper semantic understanding of the data content.

[0060] Regarding the degree of correlation C asso ( e k The calculation of ) uses a sensitivity-weighted first-order neighborhood centrality algorithm. In knowledge graphs... G KG In the middle, entity e k The risk depends not only on itself, but also on the sensitivity and connection strength of its neighboring nodes. The calculation formula is as follows:

[0061] ,

[0062] Where, N(e) k ) represents entity e k The set of first-order neighbor nodes in the graph; W sen (e j ) represents the neighbor node e j Basic sensitivity weights; Represents the edge weights between entities, reflecting the strength of the relationship; This is a normalization function to prevent numerical explosion caused by excessively high node degrees.

[0063] 2) Dynamic risk entropy calculation based on scene and environment awareness: the final risk level of data depends not only on the content but also on the dynamic constraints of the current access context. This introduces environmental trust. E trust urgency of the scenario U scene With data timeliness T fresh A dynamic risk assessment model is constructed using three dimensions.

[0064] The environmental trust score reflects the security and trustworthiness of the current data access environment. A higher score indicates a more secure environment and a lower corresponding risk entropy. This indicator is calculated by weighting three dimensions: network environment security, terminal device trustworthiness, and user behavior credibility.

[0065] ,

[0066] in, N net ( t ), D dev ( t ), Ucred ( t ) respectively represent t The network environment security, terminal device trustworthiness, and user behavior credibility at any given time are assigned discrete values. , , These are the weighting coefficients for each dimension.

[0067] Data freshness reflects the relationship between the time value and sensitivity of data over time. Real-time data typically has the highest sensitivity, while historical archived data has lower sensitivity. An exponential decay model is used for calculation.

[0068]

[0069] in, t now This is the current system access time. t gen For data i The generation timestamp, This represents the attenuation coefficient.

[0070] The scenario urgency factor is used to address the issue of "dynamic permission transfer under special circumstances." This factor is calculated based on the metadata tags of the business request.

[0071] ,

[0072] in, P level ( t () represents business priority, and the system has a preset priority mapping table. This represents the adjustment coefficient. F freq ( t This represents the number of concurrent requests for the same service tag within a short period of time, monitored through a sliding window. Define the joint risk entropy value. H risk ( i , t )as follows:

[0073] ,

[0074] in, , , These are the weighting coefficients for each dimension. For a normalized environmental trust score, the more untrustworthy the environment, the higher the risk entropy. This indicates the freshness of the data (the danger of leaking real-time data is greater than that of historical archived data). The urgency factor is adjusted using a Sigmoid function. When in high-urgency scenarios such as "emergency rescue", the denominator increases and the overall risk entropy decreases, thereby avoiding data unavailability due to excessive security control in emergency situations and demonstrating the system's adaptability.

[0075] 3) Automatic mapping and tag generation of data security levels, based on the calculated dynamic risk entropy value. H risk ( i , t Establish its relationship with discrete security levels. L k The mapping relationship is established. An adaptive threshold partitioning algorithm based on K-Means clustering is used to discretize continuous entropy values ​​into four standard security levels (L1: Public, L2: Internal, L3: Sensitive, L4: Confidential). A hierarchical mapping function is defined. F grade :

[0076] ,

[0077] in, , , This is a threshold boundary that is dynamically adjusted based on historical risk distribution. Ultimately, the system generates a metadata package containing security labels for each data object. The security level label L ( i , t This will be directly used as a global control variable, driving subsequent modules to automatically select... K - Anonymous or differential privacy algorithms, and the corresponding CP-ABE encryption strategy generated by the driver, realize an automated closed loop from "semantic awareness" to "strategy execution".

[0078] S3. Source-side adaptive privacy perturbation module

[0079] This module aims to establish a source-end adaptive perturbation mechanism driven by security level. Before data leaves the source node, it dynamically selects a de-identification strategy and adaptively adjusts the privacy budget according to its security level and the density of its spatiotemporal environment, thereby achieving "source blocking" risk control.

[0080] 1) Based on security level-based policy mapping and lightweight desensitization, the system reads the data output by module S2 with a security level L ( i , t A policy mapping mechanism is established. For low-risk data labeled L1 (public level) and L2 (internal level), k-anonymization based on generalization techniques is adopted to maximize the preservation of statistical characteristics of the data while protecting privacy. Let the original data records be...r The records after desensitization are r The transformation process is defined as follows:

[0081] ,

[0082] Where r represents a single data record of the original input; QID represents a set of quasi-identifiers, referring to a set of attributes in the data that can be combined and linked with external information to identify a specific individual, such as {date of birth, postal code, gender}; q j Represents the first in the set QID j A quasi-identifier attribute; S represents the set of sensitive attributes, referring to the core data that business analysis truly cares about; h j Represents attribute q j The currently selected generalization level; the higher the level, the more ambiguous the data. G(q) j , h j () represents a generalization function, used to map specific attribute values ​​to the first element in the domain concept tree. h j The value of the parent node of the layer.

[0083] ,

[0084] Where IL represents the amount of information lost after desensitization; M represents the total number of attributes in the quasi-identifier set QID |Range(G(q j , h j ))| represents attribute q j The size of the range of values ​​covered after generalization (e.g., if generalized to "20-30 years old", then the range size is 10); |Range(Dom j | indicates an attribute q j The total size of the interval in the global domain (e.g., if the age domain is 0-100 years old, then the total size is 100); C(r') represents the equivalence class to which record r' belongs, that is, the set of all records in the dataset that have the same value on the QID attribute; |C(r')| represents the total number of records contained in the equivalence class; k represents the privacy threshold, which requires that at least k records in the processed dataset have the same quasi-identifier characteristics, so that attackers cannot distinguish the specific target individual from these k records.

[0085] 2) Spatiotemporal density-aware adaptive allocation of privacy budget: For high-risk data marked as L3 (sensitive) and L4 (confidential), a local differential privacy protection mechanism is enabled. To address the data availability fluctuation problem caused by the fixed privacy budget in traditional LDP algorithms, a "spatiotemporal density-aware" mechanism is introduced to dynamically calculate the privacy budget for each data source node. Its core logic lies in utilizing the "crowd cover effect" of the physical environment: in dense spatiotemporal regions, natural cover is strong, and the budget can be appropriately increased to improve accuracy; in sparse regions, the budget needs to be reduced to enhance protection. First, the data source is calculated... i Real-time spatiotemporal grid density at the location :

[0086] ,

[0087] in, This represents the object density value (e.g., people per square meter) within the spatial grid s at time t. N obj ( s , t () indicates time t Located in the space grid s The total number of active objects within the area (e.g., the number of mobile phone users or vehicles currently connected to the base station in this area); Area(s) represents the physical coverage area of ​​the spatial grid s. An adaptive privacy budget adjustment function is constructed based on the density value. :

[0088] ,

[0089] in, This represents the final privacy budget dynamically allocated to data source i. A larger value indicates less injected noise and higher data availability. This represents the system's preset baseline privacy budget, indicating the default protection strength at standard density. This represents the density adjustment coefficient, used to control the weight of the impact of environmental density on the privacy budget; This represents the security level penalty factor, determined by the data security level. L Decide.

[0090] 3) Random noise injection and data distribution: For numerically sensitive data, a Laplace mechanism is employed. Let the query function be... f The global sensitivity is The generated perturbation data The calculation is as follows:

[0091] ,

[0092] in, This refers to the final, publicly released numerical value including noise. f ( D ) represents the actual result value obtained by performing a query or statistics on the original dataset D; Lap represents the random noise generation function that follows a Laplace distribution; Represents query function f Global sensitivity, which is the maximum possible change that changing a record in the dataset will cause to the query results; This represents the scale parameter of the Laplace distribution. From this equation, it can be seen that the privacy budget... The larger the value, the smaller the scale parameter, the more concentrated the generated noise distribution is around 0, and the more accurate the data. The source outputs a pre-encapsulated secure data packet. D out :

[0093] ,

[0094] in, ID mask For desensitization ID , L ( i , t (This refers to the security level.) These are the data values ​​after adding noise. This is the actual privacy budget value used (for statistically unbiased estimation correction at the receiving end).

[0095] S4. Hierarchical Policy Generation and Encrypted Storage Module

[0096] When the data stream, after undergoing source-end adaptive perturbation processing, flows into the city brain's central storage layer, it faces a fundamental contradiction between "centralized data storage" and "fine-grained access control." Furthermore, city data involves cross-departmental sharing, and a single departmental label cannot describe complex logical requirements. Therefore, this module aims to construct an encrypted storage system that separates data ownership from management rights, utilizing CP-ABE technology to encrypt data upon its arrival, and combining erasure coding technology to achieve distributed high-availability storage. The overall processing flow includes the following three parts:

[0097] 1) Automated generation of access policies based on security level mapping: To address the tedious and error-prone nature of manual policy configuration, an automated mapping mechanism from security level to logical access tree is established. Let the complete set of system attributes be... U attr (Includes attributes such as "department", "position", "rank", and "duration"). Regarding the first... i Each data record, based on its security level L ( i ) and metadata features M i Utilize strategy to generate functionsG policy Construct the access control structure tree :

[0098] ,

[0099] in, This represents the generated Boolean logic access tree. and These represent the logical "OR" and "AND" operations, respectively. j,k This is an attribute node. The mapping rule is set as follows: if L( i ) = L1 (public level), the policy tree is empty, and the data is stored in plaintext; if L( i If L(i) = L3 (Sensitive Level), departmental constraints are generated, ensuring that only personnel belonging to the designated department can decrypt the data, thus achieving horizontal isolation between departments. If L(i) = L4 (Confidential Level), multi-factor strong constraints are generated, meaning the algorithm can only successfully decrypt the data when the visitor simultaneously meets multiple attribute conditions such as identity, job level, time, and authorization status, thereby ensuring that core secrets are only visible to specific personnel in specific contexts. Through this mapping, the abstract security level is transformed into a machine-verifiable attribute logic expression, ensuring that the generation of the encryption strategy is entirely driven by the risk attributes of the data itself, without the need for manual intervention.

[0100] 2) Based on improved CP-ABE encryption, given that attribute-based encryption (CP-ABE) involves complex bilinear pairing operations, directly encrypting massive amounts of data would lead to computational performance bottlenecks. Therefore, this invention employs a hybrid encryption architecture of KEM (key encapsulation) + DEM (data encapsulation). That is, the data itself is first encrypted using an efficient symmetric key, and then the symmetric key is encrypted using the CP-ABE algorithm. The specific process is as follows: First, the system generates a random, high-strength symmetric key K. sym (e.g., a 256-bit AES key), this key is used to encrypt the plaintext data. Then, the CP-ABE algorithm is used to encrypt the K... sym Encapsulate and embed into the access policy tree In the middle, a bilinear mapping system is defined, and the generated ciphertext package is... CT i The mathematical expression is as follows:

[0101] The physical meanings of each variable are defined as follows: CT i This represents the final encrypted data packet generated and stored in the database. It contains the encrypted key, the encryption strategy structure, and necessary decryption auxiliary parameters. This represents the access policy tree structure bound to this data entry. It is stored in plaintext (e.g., ...). JSONA logical tree in a specific format (K) is used to inform the visitor what conditions need to be met for decryption, but it does not contain any key information. sym This represents the symmetric encryption key used to encrypt the original data content. e(g,g) as Represents a bilinear mapping function The result of the calculation. This represents the system's master secrets. s This represents the random secret number randomly selected by the system during each encryption process. Its purpose is to ensure that even if the same strategy is used to encrypt twice, the resulting ciphertext will be completely different, thus preventing replay attacks. This represents the masked symmetric key. Only users whose attributes meet the requirements can calculate the encryption factor and remove it, thereby restoring the encryption key. K sym h represents one of the system's public key parameters. C represents the randomization component used to assist decryption, and its value is... h s It is used to eliminate the influence of the random number s during the decryption process. This represents the set of leaf nodes in the access strategy tree. Each leaf node represents a specific attribute. and This refers to each leaf node (attribute) in the policy tree. y The generated ciphertext components contain the mathematical characteristics of the attribute, which are then used in subsequent pairing operations with the attribute components held in the user's private key to verify whether the user truly possesses the attribute.

[0102] 3) Distributed erasure coding storage and integrity verification: CP-ABE encrypted ciphertext is typically large in size and requires extremely high data reliability. To optimize storage space utilization and prevent single points of failure, (n, k) erasure coding technology is used for distributed storage. This involves cutting and encoding the encrypted object, then distributing it across different physical nodes. The matrix operations for the encoding generation process are defined as follows:

[0103] ,

[0104] in, This indicates that the original encrypted data will be... D enc Segmented k A slice of raw data. k The block contains all the information about the data. This represents the final stored data block, which usually corresponds directly to the original slice. Represents the generation through encoding calculation m One redundant check block. nThis represents the total number of data blocks to be stored, satisfying n = k + m. The system will distribute these n blocks across n different physical servers. The generator matrix, constructed using Cauchy or other methods, determines how the verification data is calculated from the original data. To prevent tampering with the ciphertext during static storage by underlying maintenance personnel, a proof-of-existence signature of the ciphertext is calculated:

[0105] ,

[0106] Among them, Sig integrity This represents the generated integrity digital signature. It is the sole credential for verifying whether the data has been tampered with. This indicates a signature operation performed using the private key of the middleware system. Indicates the encrypted data packet CT i The digest value obtained by performing a hash operation. The digest value will change drastically if the ciphertext changes by even a single bit. || represents the string concatenation operator. ID i and Timestamp These represent the unique identifier of the data and the timestamp of its entry into the database, respectively. Signing them together prevents attackers from overwriting new data with old data (replay attack) or substituting data.

[0107] S5. Multi-layered security and privacy protection module

[0108] In the deep application phase of the City Brain, data from a single department often cannot support complex cross-domain analysis needs. While traditional federated learning can achieve "data remains stationary while the model moves," existing solutions mostly employ a single encryption protocol, which is difficult to adapt to the varying security levels required by urban data—strong encryption protocols cause computational delays, while weak encryption protocols are unable to withstand high-level data reconstruction attacks. Therefore, this module aims to construct an adaptive cross-domain collaborative computing framework based on security level constraints, achieving secure alignment and joint modeling of multi-party samples without exchanging original data.

[0109] Privacy set intersection based on RSA blind signatures requires that before initiating joint modeling, participants must confirm shared sample users for data alignment, and must strictly prohibit the disclosure of non-shared samples from non-overlapping parts. A PSI protocol based on RSA blind signatures is introduced. Assume the initiator (e.g., a data platform) owns a set... C The participants have a collection S First, the initiator generates a random blinding factor. r Blinding of one's own samples:

[0110] ,

[0111] in, H For hash functions, ( e , n The initiator will use the RSA public key of each participant. T x Send to the participants. The participants use their private keys. d Sign the received blinded data, and simultaneously sign the samples you own. To perform the signature, the calculation is as follows:

[0112] ,

[0113] The participants will and Send it back to the initiator. The initiator uses the inverse of the blinding factor. right To obtain the final signature, the blinding process was performed. S x :

[0114] ,

[0115] By comparing S x and By checking if they are equal, the initiator can find the intersection. This overcomes the trust challenge in cross-domain collaboration without exposing any non-intersecting elements.

[0116] The system adaptively switches between federated protocols based on security levels. After sample alignment is completed, the system selects the data with the highest security level from the data used for modeling. L max = max( L i The gradient exchange protocol is dynamically selected. Lightweight federation (for L1 / L2 low-sensitivity data). To achieve real-time performance, gradient compression and sparsity strategies are employed. Each participant only uploads the top-k gradient parameters with the largest changes. Let the locally computed gradient vector be g, and the uploaded sparse gradient... Defined as:

[0117] ,

[0118] Strongly secure federation (for highly sensitive L3 / L4 data). To prevent gradient backpropagation attacks, it automatically switches to an MPC protocol based on additive secret sharing. (Participating parties...) k Its local gradient Split into N Each random shard is sent to other participants. The aggregation server can only receive the sum of the shards and cannot reconstruct a single gradient. The secret sharing process is as follows:

[0119] ,

[0120] in The gradient is represented by the first... m A secret fragment, P This is a large prime number field. Through this mechanism, the system achieves a dynamic optimal balance between computational efficiency and security strength.

[0121] 3) Secure Aggregation and Global Model Update: The middleware aggregation server collects encrypted gradients or gradient fragments uploaded by each participant and performs weighted aggregation to update the global model parameters. To prevent malicious participants from uploading poisoned data, a weighted aggregation function based on contribution evaluation is introduced. Let... W t These are the current global model parameters. For the first k The amount of model updates for each participant. n k Given the sample size of this participant, the model parameters for the next round... W t+1 Updated to:

[0122] ,

[0123] in, For learning rate, N total This represents the total sample size. The anomaly detection penalty coefficient is calculated based on cosine similarity.

[0124] ,

[0125] in, The average vector of all updates. This is the similarity threshold. If a participant's update direction deviates significantly from the mainstream direction (cosine similarity below the threshold),... The update is then removed by setting it to zero. This step ensures that a robust and highly accurate intelligent model of the entire city can still be built even when the original data is not visible.

[0126] S6. Selective Dense-State Query and Response Module

[0127] This module is designed with an intelligent diverter that automatically decides whether to use the "fast channel" or the "secure channel" based on the data's security level, and uses blockchain technology to permanently store query records to prevent repudiation.

[0128] 1) Intelligent query routing based on security levels: When a user initiates a query request, the system does not execute it immediately but first performs a "security check." The system parses the query statement, extracts the core fields the user wants to access, and finds the security level label generated for that field in the S2 module. The routing judgment logic is defined as follows: Let the query request be Q, and the security level of the target field be... L (data), the system's preset security threshold is L threshold (If set to L3 level). The system executes the following judgment function. Mode (Q):

[0129] ,

[0130] The plaintext high-speed mode is for publicly available data, allowing for rapid retrieval directly in memory, prioritizing maximum speed. The encrypted protection mode is for sensitive data; the system automatically intercepts requests and forces a switch to a fully homomorphic encryption engine, ensuring that the data remains undecrypted throughout the computation process.

[0131] 2) Ciphertext Retrieval and Computation in a Fully Homomorphic Environment: For highly sensitive queries that enter encrypted protection mode, the user submits the encrypted query parameter Enc(q), and the middleware directly performs homomorphic operations in the encrypted state. Let c1 = Enc(m1) and c2 = Enc(m2) be two ciphertext records stored in the database, and define homomorphic addition and homomorphic multiplication operations:

[0132] ,

[0133] in, m 1, m 2 represents plaintext data. This mechanism allows the data platform to perform mathematical calculations without knowing the data content (i.e., without decryption), and the decrypted calculation result is completely consistent with the plaintext calculation result.

[0134] End-to-end blockchain auditing and non-repudiation evidence storage are implemented to prevent data misuse or malicious tampering of query results. Consortium blockchain technology is introduced to store evidence throughout the entire query response process. At the end of the query, the system generates an audit log containing user identity, query fingerprint, and result summary, and stores it on the blockchain. The audit log structure is defined as Log. k And calculate its integrity hash H log :

[0135] Where ID_{user} is the digital identity of the querier, Hash(Q) is the hash fingerprint of the query statement, Hash(Cres) is the hash digest of the returned result, and Sign userFor users' digital signatures. To ensure the temporal order and immutability of evidence, a hash chain structure is used to anchor the current log to the blockchain block header:

[0136] ,

[0137] in, Indicates the first t The root hash value of each block, which contains the hash value of the previous block. Through this chain structure, any tampering with historical query records will cause all subsequent block hash verifications to fail, thus providing judicial-grade evidence support for tracing unauthorized queries and conducting security audits of city data.

[0138] Experimental verification

[0139] To verify the effectiveness and feasibility of the urban data adaptive privacy protection method and system based on dynamic security classification proposed in this study, an experimental platform integrating semantic dynamic classification, source-end adaptive perturbation, and dense-state collaborative computation was built in the context of comprehensive governance in key urban areas. The experimental data sources cover typical business domains such as government affairs, transportation, public health records, and community security, involving multimodal data sources including structured identity data, spatiotemporal trajectory points, business processing logs, and anonymized video keyframes. A total of 5,000 experimental samples were collected, spanning a two-week operating cycle including weekday morning and evening peak hours and holidays, exhibiting typical data heterogeneity and uneven spatiotemporal distribution characteristics. To ensure the objectivity of the experimental results and the robustness of the system, the experimental environment simulated real-world interference factors including sudden changes in application scenario status (such as switching from routine management to emergency response), drastic fluctuations in the spatiotemporal density of data sources, and a surge in cross-departmental query concurrency. This was to focus on evaluating the system's classification accuracy and the timeliness of protection strategy adjustments in complex dynamic environments. The platform is deployed based on a cloud-edge collaborative architecture simulated by a local area network. The perception layer runs a risk entropy model based on a knowledge graph, the storage layer deploys a lightweight CP-ABE policy generation engine, and the computing layer integrates adaptive differential privacy and federated learning modules to test the balance between compliance and availability of data throughout the entire process.

[0140] The comparative experiment selected the current mainstream urban data security and privacy protection schemes as the benchmark model, specifically including: (1) the traditional static access control system (RBAC), which only relies on preset static role permissions and full-disk transparent encryption for protection, lacking fine-grained dynamic classification capabilities; (2) the fixed parameter differential privacy system (Fixed-LDP), which uses a uniform and fixed privacy budget for data perturbation in all scenarios, and does not have an adaptive adjustment mechanism based on environment awareness; (3) the general fully homomorphic encryption query system (Full-FHE), which forces the full homomorphic encryption calculation for all levels of data, and does not perform traffic diversion optimization through security classification; and the adaptive protection method based on dynamic classification driven proposed in this paper. All systems were run under a unified hardware computing resource configuration, the same multi-source heterogeneous dataset, and a consistent attack simulation environment, and were divided into training-validation-test sets in a ratio of 6:2:2 to ensure the fairness and objectivity of the experimental comparison. The experimental results are as follows. Figure 2 , Figure 3 As shown in Table 1, the proposed method significantly outperforms the comparative model on most indicators, fully verifying the significant advantages of the city data adaptive privacy protection system driven by dynamic security classification in balancing security and computational efficiency.

[0141] Table 1. Comparison of data from different methods under six major indicators.

[0142] category index unit RBAC Fixed-LDP Full-FHE Method of the present invention Computational and response performance Average query response time ms 425 680 4520 315 System concurrent throughput QPS 850 720 120 1150 Privacy and security protection capabilities Sensitive information reconstruction rate % 8.45 3.20 0.05 0.08 Reasoning attack defense success rate % 15.6 78.4 99.8 98.5 Data utility and model accuracy Cross-domain model prediction accuracy % 92.1 76.5 92.1 90.8 Data utilization rate % 97 65.4 95 96.2

[0143] As shown in Figures 2 and 3 and Table 1, existing urban data protection and scheduling solutions include traditional Role-Based Access Control (RBAC), Fixed-Parameter Differential Privacy (Fixed-LDP), and Fully Homomorphic Encrypted Query (Full-FHE). While these solutions meet some requirements in certain dimensions, they struggle to achieve a perfect balance between security, efficiency, and utility, and key performance indicators still have significant shortcomings. RBAC relies on static permission lists and plaintext computation, making it ineffective against internal threats and high-concurrency surges, resulting in an average response latency of up to 425 ms and a system concurrent throughput of 850 QPS. Although Fixed-LDP enhances privacy through randomization, its rigid privacy budget allocation mechanism severely compromises data availability, causing model prediction accuracy to remain at 76.5%, failing to meet the needs of refined governance. While Full-FHE improves data utilization, its average response latency (4520 ms) remains a bottleneck due to the complex overhead of encrypted computation, making it difficult to support large-scale real-time services.

[0144] In contrast, the system proposed in this paper achieves a comprehensive performance leap driven by the synergistic effect of dynamic security classification and adaptive privacy protection. Experimental data show that this method exhibits significant advantages in key indicators of computational performance and security protection: the average query response latency is significantly reduced to 315 ms, and the system's concurrent throughput soars to 1150 QPS, significantly alleviating computational congestion; at the same time, the success rate of inference attack defense is as high as 98.5%, and the sensitive information reconstruction rate is further reduced to 0.08%, achieving high security under high concurrency. In terms of data utility and model accuracy, thanks to the spatiotemporal density-aware dynamic perturbation strategy, the cross-domain model prediction accuracy remains at 90.8%, and the data utilization rate reaches 96.2%, preserving the data value to the greatest extent while ensuring privacy, which is significantly better than the Fixed-LDP system. The above results fully verify the robustness and adaptability of this method in complex dynamic environments, proving that it has outstanding advantages in privacy protection, efficient flow, and trustworthy decision-making in the data platform of the city brain, providing solid technical support for building a secure and controllable digital city foundation.

[0145] Example 2

[0146] This embodiment provides an adaptive privacy protection system for urban data based on dynamic security classification, including:

[0147] The data acquisition module is configured to acquire data from multiple sources.

[0148] The preprocessing module is configured to perform data preprocessing based on the acquired multi-source data;

[0149] The dynamic classification module is configured to perform multi-source data semantic perception and dynamic classification based on preprocessed data, including data content semantic feature extraction based on knowledge graph, dynamic risk entropy calculation based on scene and environment perception, and automatic mapping and tag generation of data security level.

[0150] The perturbation module is configured to perform source-end adaptive privacy perturbation on the data based on the dynamic classification results, including security level-based policy mapping and lightweight desensitization, spatiotemporal density-aware privacy budget adaptive allocation, and random noise injection and data publishing.

[0151] The storage module is configured to encrypt and store the perturbation-processed data based on a hierarchical strategy, including automatic generation of access strategies based on hierarchical mapping, encryption processing based on improved CP-ABE, and distributed erasure coding storage and integrity verification.

[0152] The protection module is configured to provide multi-layered security and privacy protection for the stored data, including privacy set intersection based on RSA blind signature, adaptive switching of federated protocols based on security level, and secure aggregation and global model update.

[0153] The response module is configured to perform selective encrypted queries and responses based on the protected data.

[0154] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned method for adaptive privacy protection of urban data based on dynamic security rating.

[0155] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the aforementioned adaptive privacy protection method for urban data based on dynamic security classification.

[0156] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for adaptive privacy protection of urban data based on dynamic security classification, characterized in that, include: Acquire multi-source data; Data preprocessing is performed based on the acquired multi-source data; Multi-source data semantic perception and dynamic classification are performed based on preprocessed data, including semantic feature extraction of data content based on knowledge graph, dynamic risk entropy calculation based on scene and environment perception, and automatic mapping and tag generation of data security level. Based on the dynamic rating results, the data undergoes source-end adaptive privacy perturbation, including security level-based policy mapping and lightweight desensitization, spatiotemporal density-aware privacy budget adaptive allocation, and random noise injection and data publication. The perturbation-processed data is encrypted and stored based on a hierarchical strategy, including the automatic generation of access policies based on hierarchical mapping, encryption processing based on improved CP-ABE, and distributed erasure coding storage and integrity verification. The stored data is protected by multiple layers of security and privacy, including privacy set intersection based on RSA blind signature, adaptive switching of federated protocols based on security level, and secure aggregation and global model update. Selective encrypted query and response based on the protected data; The knowledge graph-based semantic feature extraction of data content includes extraction from standardized data objects. O i First, construct a knowledge graph for urban data security. G KG =( E , R ),in E For a collection of entities, R To establish relationships between entities, a BERT-based named entity recognition and relation extraction model is used to extract data objects. O i The feature vectors in the graph are mapped to the knowledge graph, the intrinsic sensitivity is calculated, and the data objects are defined. O i Content Sensitivity Index S content ( i ): ,in, This is the penalty coefficient for anomaly detection. e k Indicates the first extracted from the data k There are entities, I(·) is the indicator function, and W(e k C represents the pre-defined basic sensitivity weights of entities in the knowledge graph. asso ( e k ) represents an entity e k The degree of correlation in the graph, The correlation enhancement coefficient is used to determine the correlation degree C. asso ( e k The calculation of ) uses a sensitivity-weighted first-order neighborhood centrality algorithm in the knowledge graph. G KG In the middle, entity e k The risk is determined by the sensitivity and connection strength of its neighboring nodes, and is expressed as: Where, N(e) k ) represents entity e k The set of first-order neighbor nodes in the graph; W sen (e j ) represents the neighbor node e j Basic sensitivity weights; Represents the edge weights between entities. This is the normalization function; The dynamic risk entropy calculation based on scene and environment perception includes the introduction of environmental trust. E trust urgency of the scenario U scene With data timeliness T fresh A dynamic risk assessment model is constructed using three dimensions: network environment security, terminal device trustworthiness, and user behavior credibility. in, N net ( t ), D dev ( t ), U cred ( t ) respectively represent t The network environment security level, terminal device trustworthiness, and user behavior credibility at any given time are assigned discrete values. , , Here are the weighting coefficients for each dimension; an exponential decay model is used to calculate data freshness. in, t now This is the current system access time. t gen For data i The generation timestamp, Represents the attenuation coefficient; the urgency factor of the scenario is mapped and calculated based on the metadata tags of the business request: in, P level ( t () represents business priority. Represents the adjustment coefficient. F freq ( t This represents monitoring the concurrent request volume of the same business tag within a short period of time through a sliding window, defining a joint risk entropy value. H risk ( i , t )as follows: ， in, , , These are the weighting coefficients for each dimension. For normalized environmental trust scores, Indicates the freshness of the data. As a factor of urgency in the scenario; The security level-based policy mapping and lightweight de-identification include, based on the output data security level L ( i , t A policy mapping mechanism is established. For low-risk data labeled as L1 and L2, a generalized k-anonymization process is adopted to maximize the preservation of statistical characteristics of the data while protecting privacy. The spatiotemporal density-aware privacy budget adaptive allocation includes a spatiotemporal density-aware mechanism to dynamically calculate the privacy budget for each data source node, addressing the data availability fluctuation problem caused by the fixed privacy budget in the traditional LDP algorithm for high-risk data marked as sensitive and confidential.

2. The adaptive privacy protection method for urban data based on dynamic security classification as described in claim 1, characterized in that, The data preprocessing based on the acquired multi-source data includes adopting a unified data access framework for data streams from different sources, and implementing multi-protocol parsing of the original input dataset through an interface adaptation layer. D for: ,in, d i Indicates the first i One original data record, s i Indicates the data source identifier. f i Indicates the first i The original set of fields contained in each data record. t i Using timestamps, a unified data standardization mapping function is defined for both structured and semi-structured data to obtain standardized feature values: in, Indicates the first i The first data record j The original values ​​of each field, and These represent the mean and standard deviation of the field in the global data, respectively. For unstructured text, key fields are extracted and entities are standardized through natural language processing, and semantic tags with unified names are extracted to achieve semantic-level fusion. For video streams or image data, frame extraction and time indexing are used to transform unstructured information into structured descriptions, ultimately forming data objects in a unified format. in, ID i Indicates the data object number. Type i Represents data type, Feature i This represents a vectorized feature description generated by an encoding network or feature extraction model. Time i Indicates the time index of the data. Location i This represents the spatial location information of the data; to achieve spatiotemporal consistency across data sources, a synchronization algorithm based on time windows and a spatial coordinate transformation mechanism are introduced, employing a linear interpolation method based on the target time point. t * Estimate the data from two adjacent time points: ， in, x t and x t+1 They are time points respectively t and t The data value at +1, t * Indicates the target synchronization time; regarding spatial alignment, for spatial data from different coordinate systems, a two-dimensional affine transformation model is used to achieve coordinate unification: ， where (x, y) and (x', y') are the geographic location points in the original and target coordinate system, respectively, and matrix parameters a 11 , a 12 , a 21 , a 22 represent rotation and scaling parameters, t x , t y represent translation vectors, used to adjust spatial offsets, and through this transformation, spatial unification of urban multi-department system data is achieved.

3. The adaptive privacy protection method for urban data based on dynamic security classification as described in claim 2, characterized in that, The security level-based policy mapping and lightweight de-identification include, based on the output data security level L ( i , t A policy mapping mechanism is established. For low-risk data labeled as L1 and L2, to maximize the preservation of statistical characteristics while protecting privacy, a generalized k-anonymity processing method is adopted. Let the original data records be... r The records after desensitization are r The transformation process is defined as follows: Where r represents a single data record of the original input; QID represents the set of quasi-identifiers, q j Represents the first in the set QID j A quasi-identifier attribute; S represents the set of sensitive attributes; h j Represents attribute q j The currently selected generalization level, G(q) j , h j () represents a generalization function, used to map specific attribute values ​​to the first element in the domain concept tree. h j The parent node value of the layer, ， Where IL represents the amount of information lost after desensitization; M represents the total number of attributes in the quasi-identifier set QID, |Range(G(q j , h j ))| represents attribute q j The size of the range of values ​​covered after generalization; |Range(Dom j | indicates an attribute q j The total interval size in the global domain; C(r') represents the equivalence class to which record r' belongs; |C(r')| represents the total number of records contained in the equivalence class; k represents the privacy threshold; the spatiotemporal density-aware privacy budget adaptive allocation includes, for high-risk data marked as sensitive and confidential, addressing the data availability fluctuation problem caused by the fixed privacy budget in the traditional LDP algorithm, introducing a spatiotemporal density-aware mechanism to dynamically calculate the privacy budget of each data source node. First, the data source... i Real-time spatiotemporal grid density at the location : in, This represents the object density value within the spatial grid s at time t; N obj ( s , t () indicates time t Located in the space grid s The total number of active objects within the area; Area(s) represents the physical coverage area of ​​the spatial grid s, and an adaptive privacy budget adjustment function is constructed based on the density value. : in, This represents the final privacy budget dynamically allocated to data source i; This represents the system's preset baseline privacy budget value; Indicates the density adjustment coefficient; This represents the security level penalty factor, determined by the data security level. L Decide.

4. The adaptive privacy protection method for urban data based on dynamic security classification as described in claim 3, characterized in that, The automated generation of access policies based on security level mapping includes establishing an automated mapping mechanism from security level to logical access tree, assuming the complete set of system attributes is... U attr Regarding the first i Each data record, based on its security level L ( i ) and metadata features M i Generate functions using strategies G policy Construct an access control structure tree : in, This represents the generated Boolean logic access tree. and These represent the logical "OR" and "AND" operations, respectively. j,k For attribute nodes, the abstract security level is transformed into a machine-verifiable attribute logical expression through mapping; the improved CP-ABE encryption processing includes a hybrid encryption architecture using KEM key encapsulation and DEM data encapsulation, first generating a random high-strength symmetric key K. sym The key is used to encrypt the plaintext data, and then the CP-ABE algorithm is used to encrypt the K. sym Encapsulate and embed into the access policy tree In the definition of a bilinear mapping system, the generated ciphertext package is packaged as follows. CT i Mathematical expression: in, CT i This refers to the encrypted data packets that are ultimately generated and stored in the database. This represents the access policy tree structure bound to this data, K. sym Let e(g,g) represent the symmetric encryption key used to encrypt the original data content. as Represents a bilinear mapping function The result of the calculation This represents the system's master secret parameter. s This represents the secret random number randomly selected by the system during each encryption process. Let C represent the masked symmetric key, and let C represent the randomization component used to assist decryption. This represents the set of leaf nodes in the access strategy tree. and This represents the ciphertext component generated for each leaf node in the policy tree.

5. The adaptive privacy protection method for urban data based on dynamic security classification as described in claim 4, characterized in that, The privacy set intersection method based on RSA blind signatures includes introducing a PSI protocol based on RSA blind signatures, assuming the initiator owns the set. C The participants have a collection S First, the initiator generates a random blinding factor. r Blinding of one's own samples: in, H For hash functions, ( e , n The initiator will use the RSA public key of the participating party. T x Send to the participants, who then use their private keys. d Sign the received blinded data, and simultaneously sign the samples you own. To perform the signature, the calculation is as follows: The participants will and It is sent back to the initiator, who uses the inverse of the blinding factor. right To obtain the final signature, the blinding process was performed. S x : By comparing S x and By checking if they are equal, the initiator can find the intersection. This overcomes the trust challenge of cross-domain collaboration without exposing any non-intersecting elements.

6. The adaptive privacy protection method for urban data based on dynamic security classification as described in claim 5, characterized in that, The adaptive switching of the federated protocol based on security levels includes, after completing sample alignment, switching based on the highest security level of the data involved in modeling. L max = max( L i The gradient exchange protocol is dynamically selected, and gradient compression and sparsity strategies are adopted. Let the locally computed gradient vector be g, and the uploaded sparse gradient... Defined as: All participating parties k Its local gradient Split into N A random shard is sent to other participants, and the secret sharing process is as follows: in The gradient is represented by the first... m A secret fragment, P For the large prime number field; the secure aggregation and global model update include introducing a weighted aggregation function based on contribution evaluation, assuming... W t These are the current global model parameters. For the first k The amount of model updates for each participant. n k Given the sample size of this participant, the model parameters for the next round... W t+1 Updated to: in, For learning rate, N total The total sample size is... The anomaly detection penalty coefficient is calculated based on cosine similarity. ， in, The average vector of all updates. This is the similarity threshold.

7. The adaptive privacy protection method for urban data based on dynamic security classification as described in claim 6, characterized in that, The selective encrypted query and response based on the protected data includes extracting the core fields the user wants to access by parsing the query statement according to the user's query request, and finding the security level label generated by the field. For highly sensitive queries that enter encrypted protection mode, the user submits the encrypted query parameter Enc(q), and the middleware directly performs homomorphic operations in the encrypted state. Let c1 = Enc(m1) and c2 = Enc(m2) be two encrypted records stored in the database, and define homomorphic addition and homomorphic multiplication operations: ， in, m 1, m 2 represents plaintext data. To prevent data misuse or malicious tampering of query results, consortium blockchain technology is introduced to preserve evidence throughout the entire query response process. At the end of the query, an audit log containing user identity, query fingerprint, and result summary is generated and stored on the blockchain. The audit log structure is defined as Log. k And calculate its integrity hash H log : Where ID_{user} is the digital identity of the querier, Hash(Q) is the hash fingerprint of the query statement, Hash(Cres) is the hash digest of the returned result, and Sign user Digital signature for users.

8. A city data adaptive privacy protection system based on dynamic security classification, executing the city data adaptive privacy protection method based on dynamic security classification as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire data from multiple sources. The preprocessing module is configured to perform data preprocessing based on the acquired multi-source data; The dynamic classification module is configured to perform multi-source data semantic perception and dynamic classification based on preprocessed data, including data content semantic feature extraction based on knowledge graph, dynamic risk entropy calculation based on scene and environment perception, and automatic mapping and tag generation of data security level. The perturbation module is configured to perform source-end adaptive privacy perturbation on the data based on the dynamic classification results, including security level-based policy mapping and lightweight desensitization, spatiotemporal density-aware privacy budget adaptive allocation, and random noise injection and data publishing. The storage module is configured to encrypt and store the perturbation-processed data based on a hierarchical strategy, including automatic generation of access strategies based on hierarchical mapping, improved CP-ABE encryption processing, and distributed erasure coding storage and integrity verification. The protection module is configured to provide multi-layered security and privacy protection for the stored data, including privacy set intersection based on RSA blind signature, adaptive switching of federated protocols based on security level, and secure aggregation and global model update. The response module is configured to perform selective encrypted queries and responses based on the protected data.

Images