Distributed power data asset management privacy protection encryption architecture design method

Abstract

The invention provides a distributed power data asset management privacy protection encryption architecture design method, which comprises the steps of generating a watermark with robustness and imperceptibility by adopting a content-based digital watermarking algorithm according to statistical characteristics of power data assets, and storing a watermark embedding process and an extraction result through a blockchain; a graph-based data model is adopted, the electric power data assets carrying the watermarks serve as nodes, the data flow relation serves as edges, and an electric power data asset consanguinity graph is constructed; for distributed storage of the blood relationship map, a P2P network topology is constructed by adopting a block chain and a distributed hash table, and blood relationship map data is stored in each node of an alliance chain in a distributed manner; and verifying the digital watermark information embedded in all circulation path nodes through which the power data assets pass in the blood relationship map, calculating the similarity between the extracted watermark and the original watermark, calculating the distortion rate, and judging whether the power data assets are tampered or not.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a design method for a distributed power data asset management privacy protection encryption architecture. Background Art

[0002] In the process of power data asset management, since the circulation paths of power data assets are complex and diverse, involving many participants, how to effectively manage the circulation path information of power data assets and ensure the privacy and security of power data assets during the circulation process has become a technical problem that needs to be solved urgently. At present, the commonly used power data asset management method is to identify and track power data assets through digital watermarking technology, but this method can only track the flow of power data assets and cannot comprehensively manage the circulation path information of power data assets. In addition, most of the existing digital watermarking technologies use unified watermark information, which is difficult to adapt to the dynamic changes of power data assets, resulting in power data assets being easily tampered with and forged during the circulation process, and privacy and security are difficult to guarantee. Therefore, there is an urgent need for a power data asset lineage analysis and privacy protection technology that can dynamically generate unique watermarks, distribute the management of power data asset circulation path information, and adapt to dynamic changes, so as to solve the privacy and security issues in the circulation process of power data assets and realize effective distributed management of the entire life cycle of power data assets. Summary of the invention

[0003] The present invention provides a distributed power data asset management privacy protection encryption architecture design method, which mainly includes:

[0004] According to the statistical characteristics of power data assets, a content-based digital watermark algorithm is used to generate a robust and imperceptible watermark, and the watermark embedding process and extraction results are stored on the blockchain.

[0005] Aiming at watermark embedding and extraction in distributed environment, a watermark protocol for secure multi-party computing is designed. By using secret sharing and blind watermark mechanism, privacy-preserving embedding and blind extraction of watermarks of power data assets are performed to obtain power data assets carrying watermarks.

[0006] A graph-based data model is adopted, with watermarked power data assets as nodes and data flow relationships as edges, to construct a power data asset lineage graph;

[0007] For the distributed storage of bloodline graphs, blockchain and distributed hash tables are used to build a P2P network topology, and the bloodline graph data is distributed and stored in each node of the alliance chain;

[0008] Verify the digital watermark information embedded in all the circulation path nodes of the power data assets in the lineage map, calculate the similarity between the extracted watermark and the original watermark, calculate the distortion rate, and determine whether the power data assets have been tampered with;

[0009] When the watermark extracted from the power data node in the bloodline graph is distorted, access control strategies of different granularities are designed to generate attribute encryption keys associated with user attributes, and fine-grained data authorization and permission management are performed. At the same time, a proxy re-encryption mechanism is introduced to convert attribute encrypted ciphertext into a form that can be decrypted by different users, and Laplace noise is injected into the ciphertext conversion process through differential privacy.

[0010] Based on the federated learning framework, data asset nodes in the bloodline graph are used as samples, and collaborative training is carried out in a distributed environment. The data asset ciphertext after differential privacy injection of Laplace noise is modeled and analyzed, and secure multi-party computation of data asset ciphertext is performed to conduct fusion analysis of power data assets in a confidential state.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a distributed power data asset management privacy protection encryption architecture design method. Aiming at the protection and management issues of power data assets, the present invention realizes data robustness protection and privacy security management through the following key technical links:

[0013] First, a content-based digital watermark algorithm is used to combine the statistical characteristics of power data assets to generate a robust and imperceptible watermark. The present invention records and verifies the watermark embedding and extraction process and results through blockchain technology, ensuring the immutability and traceability of the watermark information.

[0014] Secondly, for the watermark embedding and extraction of power data assets in a distributed environment, the present invention designs a watermark protocol for secure multi-party computing. Through secret sharing and blind watermarking mechanisms, the privacy-preserving embedding and blind extraction of power data asset watermarks are achieved, ensuring the privacy security of data during processing.

[0015] Thirdly, by adopting a graph-based data model, the present invention constructs a lineage graph of power data assets with watermarks as nodes and data flow relationships as edges. By integrating the P2P network topology constructed by blockchain and distributed hash tables into the distributed storage solution of the lineage graph, distributed storage and efficient management of data are achieved.

[0016] In addition, the present invention also introduces fine-grained access control strategies and attribute encryption technology. When the power data asset node in the lineage graph detects watermark distortion, it can generate attribute encryption keys associated with user attributes to perform data authorization and permission management. Through differential privacy technology and proxy re-encryption mechanism, the security of data in the sharing and conversion process is further enhanced.

[0017] Finally, based on the federated learning framework, the present invention conducts collaborative training and model analysis on the differentially privacy protected power data asset ciphertext, realizes the secure multi-party computation and fusion analysis of the data asset ciphertext, and ensures the privacy protection and security of the data during the analysis and processing process.

[0018] In general, the present invention effectively solves the security management and privacy protection issues of power data assets in a distributed environment, improves data security and management efficiency, and protects the integrity and reliability of data assets. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a method for designing a privacy-preserving encryption architecture for distributed power data asset management according to the present invention.

[0020] Figure 2 A schematic diagram of a distributed power data asset management privacy protection encryption architecture design method of the present invention.

[0021] Figure 3 This is another schematic diagram of a distributed power data asset management privacy protection encryption architecture design method of the present invention. DETAILED DESCRIPTION

[0022] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0023] like Figure 1-3 In this embodiment, a distributed power data asset management privacy protection encryption architecture design method may specifically include:

[0024] S101. Based on the statistical characteristics of power data assets, a content-based digital watermarking algorithm is used to generate a robust and imperceptible watermark, and the watermark embedding process and extraction results are stored on the blockchain.

[0025] Time series analysis and spectrum analysis are performed on the power data assets to obtain the statistical feature vector of the power data assets, and the statistical feature vector includes mean, variance, autocorrelation coefficient and power spectrum density. According to the statistical feature vector, a digital watermark algorithm is designed by discrete cosine transform, and the watermark embedding parameter is obtained by adjusting the watermark embedding strength coefficient. The power data assets are multi-scale decomposed by discrete wavelet transform to obtain low-frequency sub-bands. An embedding position is selected in the low-frequency sub-band, and the watermark information is encoded by spread spectrum watermark technology to obtain the encoded watermark information. The watermark embedding parameter and the encoded watermark information are hashed to generate a unique identifier. A block data structure is constructed in combination with the timestamp information, and the block data structure is added to the blockchain network by using the proof of stake consensus mechanism. A watermark extraction verification scheme is designed based on the zk-SNARKs zero-knowledge proof protocol, and the verification logic is deployed in the blockchain smart contract. If a watermark extraction request is received, the verification logic is executed. The verification logic includes: obtaining the unique identifier from the blockchain network, performing a hash operation on the watermark information in the watermark extraction request, and obtaining a hash result. It is determined whether the hash result matches the unique identifier, and if so, it is determined that the watermark extraction request is valid.

[0026] Specifically, time series analysis and spectrum analysis are used to extract features of power data assets to obtain statistical feature vectors of the data, including mean, variance, autocorrelation coefficient and power spectral density. According to the extracted feature vectors, a digital watermark algorithm based on discrete cosine transform is designed, and the robustness and imperceptibility of the watermark are balanced by adjusting the watermark embedding strength coefficient. The power data is decomposed at multiple scales using discrete wavelet transform, and the embedding position is selected in the low-frequency subband. The watermark information is encoded using spread spectrum watermarking technology to enhance the robustness of the watermark. Dequantization and inverse wavelet transform operations are performed to extract watermark information from the watermarked power data. The watermark embedding and extraction process is hashed to generate a unique identifier, and the block data structure is constructed in combination with the timestamp information. The block is added to the blockchain network using the proof-of-stake consensus mechanism to achieve tamper-proof storage of the watermark processing process. A watermark extraction verification scheme is designed based on the zk-SNARKs zero-knowledge proof protocol, and the verification logic is deployed in the blockchain smart contract, including defining the verification function, setting the verification parameters and writing the verification rules. The intermediate data in the watermark extraction process is encrypted through lattice cryptography homomorphic encryption technology to protect data privacy. The smart contract receives the encrypted extraction result, compares it with the hash value stored on the chain, verifies the integrity of the watermark, and returns the verification result.

[0027] The power data assets were analyzed by time series and spectrum. The mean of the power consumption data at 15-minute intervals was calculated to be 500kWh, the variance was 100kWh, the autocorrelation coefficient was 0.85, and the power spectrum density peak appeared at 0.2Hz. Based on these characteristics, a discrete cosine transform watermark algorithm was designed, and the intermediate frequency coefficient was selected for watermark embedding, and the embedding strength coefficient was set to 0.05. The power data was decomposed using a 4-level discrete wavelet transform, and 50 coefficients were selected as embedding positions in the LL4 subband. The watermark information was encoded using a direct sequence spread spectrum with a spreading factor of 16 to generate a 128-bit watermark sequence. The dequantization operation was performed, and the threshold was set to 0.03. The extracted watermark sequence was correlated and detected. The watermark was determined to exist when the correlation peak was greater than 0.7. The SHA-256 algorithm was used to calculate the hash value of the watermark embedding process to generate a 32-byte unique identifier. The block data structure was constructed in conjunction with the Unix timestamp 1699000000, and the block size was 1MB. The proof-of-stake consensus mechanism is adopted, the verification node holds 100 tokens, and the block time is 10 seconds. The verification scheme is designed using the zk-SNARKs protocol to generate a 128-byte proof. The smart contract defines the verifyWatermark function and sets the minimum verification threshold to 0.8. The BGV homomorphic encryption scheme in lattice cryptography is used with a key length of 2048 bits to encrypt the intermediate data in the watermark extraction process. The smart contract receives the encryption result and compares it with the stored hash value. When the similarity is greater than 0.9, the verification is passed, and the Boolean value true is returned to indicate that the integrity of the watermark is confirmed.

[0028] S102. Aiming at watermark embedding and extraction in a distributed environment, a watermark protocol for secure multi-party computing is designed. Secret sharing and blind watermarking mechanisms are used to perform privacy-preserving embedding and blind extraction of watermarks for power data assets, and obtain power data assets carrying watermarks.

[0029] The watermark information is segmented by Shamir threshold secret sharing scheme, the threshold value is set to t, n watermark information fragments are obtained, and distributed to n distributed nodes. According to the time series characteristics of the watermark information fragments and the power data assets, the discrete cosine transform domain is selected as the watermark embedding space, and the transform coefficients are encrypted by the Paillier homomorphic encryption algorithm. The watermark embedding operation is performed in the encrypted domain, the intermediate frequency coefficient is selected as the embedding position, the coefficient amplitude is adjusted according to the watermark information fragments, and the coefficient modification of the encrypted domain is realized by using the additive property of homomorphic encryption. The watermark extraction protocol is designed through the blind watermark mechanism, the watermark information is extracted in the encrypted domain, and the Schnorr zero-knowledge proof protocol is used to verify the existence of the watermark. If the verification is successful, the (t, n) threshold RSA signature scheme is used to verify the final result, generate an RSA key pair, obtain n private key fragments, and distribute them to the n distributed nodes; obtain partial signatures from t nodes in the n distributed nodes; combine the t partial signatures to obtain a complete signature; and verify the validity of the complete signature.

[0030] Specifically, the Shamir threshold secret sharing scheme is used to divide the watermark information, and the threshold value is set to t. The watermark information is divided into n parts and distributed to n distributed nodes. After any t nodes receive the shares, the complete watermark information is reconstructed by the Lagrange interpolation method. The reconstructed watermark information is preprocessed, including normalization and encoding, to make it suitable for embedding in the discrete cosine transform domain. According to the time series characteristics of the power data asset, the discrete cosine transform domain is selected as the watermark embedding space, and the transform coefficients are encrypted using the Paillier homomorphic encryption algorithm. The watermark embedding operation is performed in the encrypted domain, and the specific steps include: selecting the intermediate frequency coefficient as the embedding position, adjusting the coefficient amplitude according to the watermark information, and using the additive property of homomorphic encryption to realize the coefficient modification in the encrypted domain. The watermark extraction protocol is designed through the blind watermark mechanism, and the Schnorr zero-knowledge proof protocol is used to verify the existence of the watermark without knowing the original power data. During the verification process, the prover constructs the commitment value, the verifier sends a random challenge, the prover calculates the response based on the challenge, and the verifier checks the validity of the response. The watermark information is extracted from the encrypted domain and collaboratively verified in a distributed environment. Each node compares the extracted result with the original watermark information. The secure multi-party computing protocol is used to perform collaborative computing in a distributed environment, and each node signs the local computing result. The (t,n) threshold RSA signature scheme is used to verify the final result. The specific steps include: generating an RSA key pair, dividing the private key into n parts, and distributing it to n nodes; t nodes cooperate to generate partial signatures; combining t partial signatures to obtain a complete signature; and verifying the validity of the signature. Finally, the power data asset carrying the watermark and maintaining the original load curve characteristics is obtained. Taking 128-bit watermark information as an example, the Shamir (3,5) threshold secret sharing scheme is used for segmentation, and the polynomial f(x) = s + a1x + a2x^2 on the finite field GF(2^8) is selected, where s is the watermark information and a1 and a2 are randomly selected coefficients. The function values ​​of 5 points are calculated to obtain 5 secret shares. After receiving the shares, the 3 nodes use the Lagrange interpolation method to reconstruct the watermark information. The reconstructed watermark information is normalized, mapped to the interval [-1,1], and encoded using Hamming code (7,4) to obtain a 224-bit encoded watermark sequence. The power load data at 15-minute intervals is selected and a 64-point discrete cosine transform is performed on it. The Paillier homomorphic encryption algorithm is used to select a 1024-bit key length to encrypt the transform coefficients. In the encryption domain, the coefficients of the intermediate frequency 16-31 are selected as the watermark embedding position, the coefficient amplitude is adjusted according to the watermark bit value, and the embedding strength is set to 0.05. The Schnorr zero-knowledge proof protocol is used to verify the existence of the watermark, and the prime number p=2^256-189 and the generator g=2 are selected. The prover selects a random number r and calculates the commitment value T=g^rmodp. The verifier sends a random challenge c∈[0,2^128-1].The prover calculates the response s = (r + cx) mod (p-1), where x is the private key. The verifier checks whether g^s = T·y^c mod p holds, where y is the public key. The watermark information is extracted from the encrypted domain, and the relevant detection threshold is set to 0.7. Collaborative verification is performed on 5 distributed nodes, and each node compares the Hamming distance between the extracted result and the original watermark, and the threshold is set to 20. The (3,5) threshold RSA signature scheme is adopted, and the 3072-bit RSA modulus n is selected to divide the private key d into 5 parts. The three nodes respectively calculate the partial signature si = m^(d_i) mod n, where m is the message hash value. The three partial signatures are combined to obtain the complete signature s = s1·s2·s3modn. Finally, the power data asset carrying the watermark is obtained. By comparing the peak-to-valley difference, load factor and other indicators between the original data and the watermark data, it is ensured that the original load curve characteristics are maintained.

[0031] S103. Using a graph-based data model, with watermarked power data assets as nodes and data flow relationships as edges, a power data asset lineage graph is constructed.

[0032] Define the attribute fields of the nodes in the graph data model, and the attribute fields include asset identification, data type, timestamp, spatial location and watermark information. Use Neo4j attribute graph database to store the node information, and use Cypher language to create the node structure. Construct data flow relationship edges according to the node information, and calculate the weight value of the relationship edge, and the weight value is determined by the data flow frequency, data volume and data importance. Use the depth-first search algorithm to traverse the graph constructed by the relationship edge to identify the blood relationship between data assets. Use the Louvain community detection algorithm to optimize the blood relationship graph structure, merge redundant nodes and edges, and obtain an optimized blood relationship map. Serialize the optimized blood relationship map into JSON format and store it in the Neo4j database. Read the serialized blood relationship map from the Neo4j database and deserialize it into a graph structure. Use the D3.js visualization library to implement a graphical interface and convert the graph structure data into a force-directed graph layout. For the nodes carrying watermarks in the force-directed graph layout, a watermark verification function is called to check the integrity of the watermark, and different colors are used in the graphical interface to mark nodes with normal and abnormal watermark states.

[0033] Specifically, according to the characteristics of power data assets, the attribute fields of nodes in the graph data model are defined, including asset identification, data type, timestamp, spatial location and watermark information, and the Neo4j property graph database is used to store node information. When creating a node in Neo4j, the Cypher language is used to define the node structure, and the primary key index is set to improve query efficiency. By analyzing the business associations and data processing processes between power data assets, the data flow relationship edges are constructed and the weight values ​​of the edges are calculated. The weight value calculation formula is: W = α*F+β*V+γ*I, where F is the data flow frequency, V is the data volume, I is the data importance, and α, β, and γ are weight coefficients. The constructed graph is traversed using the depth-first search algorithm to identify the blood relationship between data assets and construct the blood relationship map of power data assets. The Louvain community detection algorithm is used to optimize the graph structure, merge redundant nodes and edges, and improve the readability of the graph. The optimized blood relationship map is serialized into JSON format and stored in the Neo4j database. Neo4j's graph database indexing technology is used to improve the efficiency of graph queries and realize graph-based data traceability. The serialized lineage graph is read from the Neo4j database and deserialized into a graph structure. The D3.js visualization library is used to implement a graphical interface, converting the graph structure data into a force-directed graph layout, supporting node dragging, zooming, and hovering to display detailed information. During the graph query process, for nodes carrying watermarks, the watermark verification function is called to check the integrity of the watermark, and different colors are used in the visualization interface to mark nodes with normal and abnormal watermark status. Taking the distribution network data of a certain power company as an example, the node attributes in the graph data model are defined as follows: asset ID (string, primary key), data type (enumeration: voltage, current, power, etc.), timestamp (Unix timestamp), spatial location (latitude and longitude coordinates) and watermark information (128-bit binary string). The Neo4j database is used to store node information, and the Cypher statement for creating a node is:

[0034] CREATE(n:PowerAsset{assetId:'001′,dataType:′voltage′,timestamp:1625097600,location:′39.9042,116.4074',watermark:'1010...1101'}). Set asset ID as primary key index: CREATEINDEXON:PowerAsset(assetId). Analyze data flow relationships, such as voltage data transmission from substation to distribution line, and build edge relationships. In edge weight calculation, set α=0.5, β=0.3, γ=0.2, F takes values ​​of 0-100 times / day, V takes values ​​of 0-1TB, and I takes values ​​of 1-10. For example, if F=50, V=0.5TB, and I=8 for an edge, then W=0.550+0.30.5+0.2*8=26.15. Use depth-first search to traverse the graph, starting with the main substation node, and recursively visit adjacent nodes until the leaf node. Louvain algorithm optimizes the graph structure: Initially, each node is regarded as a community, iteratively calculates the modularity gain between nodes, and merges the communities with the largest modularity gain until the modularity no longer increases. The optimized graph is serialized into JSON: {"nodes":[{"id":"001","type":"voltage",...}],"edges":[{"source":"001","target":"002","weight":26.15}]}. Neo4j query uses indexes to improve efficiency: MATCH(n:PowerAsset)WHEREn.assetId='001'RETURNn. Use D3.js for visualization, and set the force-directed graph parameters: the number of mechanical simulation iterations is 300, the repulsion between nodes is -30, and the length of the connection line is 100. The node size is mapped to the range of 10-30 pixels according to the number of connections. The color is red for abnormal watermarks and green for normal. The watermark verification function compares the stored watermark with the extracted watermark, and determines it as complete if the Hamming distance is less than 5. Through these specific implementation methods, the construction, optimization, storage, query and visualization of the power data asset lineage map are realized.

[0035] S104. For the distributed storage of bloodline graphs, blockchain and distributed hash tables are used to build a P2P network topology, and the bloodline graph data is distributed and stored in each node of the alliance chain.

[0036] The Chord algorithm is used to construct a distributed hash table, and the bloodline map data is serialized into a key-value pair format, where the key is the node identifier and the value is the node attribute and adjacency relationship. According to the key-value pair format data, the mapping relationship between the data and the alliance chain node is calculated by the consistent hash function. The blockchain consensus mechanism is designed based on the practical Byzantine fault tolerance algorithm to synchronize the update operation of the bloodline map data between the alliance chain nodes. The consensus mechanism includes five stages: master node election, request broadcast, pre-preparation, preparation and submission, and the data validity is verified through the voting mechanism. The bloodline map data transmitted between nodes is encrypted using the RSA asymmetric encryption algorithm, and the node identity authentication is realized in combination with the zk-SNARK zero-knowledge proof protocol. In the zero-knowledge proof protocol, the prover generates a proof containing the node identity information, and the verifier verifies the validity of the proof without obtaining the specific identity information. A distributed index based on the graph structure is constructed, and the topological characteristics of the bloodline map are used to establish an inverted index of the relationship between nodes. A distributed graph query algorithm is designed to decompose the complex bloodline map query task into sub-queries. In the distributed graph query algorithm, the query plan generator formulates the query path according to the topological structure characteristics of the graph. The subqueries are distributed to the relevant alliance chain nodes through the P2P network, and the query operations are executed in parallel. The MapReduce framework is used to summarize the query results. The Map stage processes the subquery results, and the Reduce stage merges the final query results.

[0037] Specifically, the Chord algorithm is used to construct a distributed hash table, and the bloodline map data is serialized into a key-value pair format, where the key is the node ID and the value is the node attribute and adjacency relationship. The mapping relationship between data and alliance chain nodes is calculated through the consistent hash function to achieve uniform distribution and rapid positioning of data. The blockchain consensus mechanism is designed based on the practical Byzantine fault tolerance algorithm to synchronize the update operation of bloodline map data between alliance chain nodes. The consensus process includes five stages: master node election, request broadcast, pre-preparation, preparation and submission. The validity of the data is verified through the voting mechanism to ensure data consistency. The RSA asymmetric encryption algorithm is used to encrypt the bloodline map data transmitted between nodes, and the zk-SNARK zero-knowledge proof protocol is combined to realize node identity authentication and data access control. In zk-SNARK, the prover generates a proof containing node identity information, and the verifier verifies the validity of the proof without obtaining specific identity information. A distributed index based on the graph structure is constructed, and the topological characteristics of the bloodline map are used to establish an inverted index of the relationship between nodes to improve data retrieval efficiency. A distributed graph query algorithm is designed to decompose complex bloodline map query tasks into subqueries. The query plan generator formulates the optimal query path based on the topological structure characteristics of the graph. Subqueries are distributed to relevant alliance chain nodes through the P2P network, and query operations are executed in parallel. The MapReduce framework is used to summarize query results. The Map stage processes subquery results, and the Reduce stage merges the final query results to improve query efficiency.

[0038] Taking the lineage graph of a power company's distribution network as an example, the Chord algorithm is used to build a distributed hash table containing 1024 nodes. The lineage graph data is serialized into key-value pairs, such as {key: "transformer_001", value: {type: "transformer", voltage: 10kV, connected_to: ["line_002", "line_003"]}}. The SHA-256 hash function is used to calculate the data and node mapping to ensure that the data is evenly distributed. In the practical Byzantine fault tolerance algorithm, the fault tolerance threshold f = 3 is set, that is, the system allows up to 3 nodes to fail. The master node election adopts a rotation system, and the master node is replaced every 100 blocks. During the consensus process, the timeout time is set to 5 seconds in the request broadcast stage, and the thresholds for the pre-preparation and preparation stages are 2f+1 = 7 nodes to reach a consensus. RSA encryption uses a 2048-bit key to ensure data transmission security. In the zk-SNARK protocol, the upper limit of proof generation time is set to 200 milliseconds, and the upper limit of verification time is set to 50 milliseconds to ensure the efficiency of identity authentication. A distributed index based on the graph structure is constructed, and for each node, an inverted index of adjacency relationships within a range of up to 10 hops is established. In the distributed graph query algorithm, the query plan generator uses a dynamic programming algorithm to generate the optimal query path considering the node degree and edge weight. The subquery parallelism is set to 16, that is, the query is executed on 16 nodes at the same time. In the MapReduce framework, the timeout of the Map task is 30 seconds, and the timeout of the Reduce task is 60 seconds. Through these specific implementation methods, the distributed storage, secure transmission, efficient query and consistency maintenance of the bloodline graph of power data assets are realized.

[0039] S105: Verify the digital watermark information embedded in all the circulation path nodes that the power data asset passes through in the lineage map. By calculating the similarity between the extracted watermark and the original watermark and calculating the distortion rate, it is determined whether the power data asset has been tampered with.

[0040] A depth-first search algorithm is used to traverse the lineage graph, identify all the circulation path nodes through which the power data asset passes, and store the node information in an ordered list. Watermark extraction parameters are selected according to the node type, and the watermark extraction parameters include wavelet transform levels and subband selection. The power data is multi-scaled decomposed using a 3-level discrete wavelet transform, and the embedded digital watermark information is extracted in the LL3 subband. A sliding window method is used for correlation detection, and the window size of the sliding window method is set to 1 / 4 of the watermark length, and the step size is 1. The watermark sequence is restored by calculating the local correlation peak. If the extraction fails, the BCH error correction code is used to repair the watermark. The normalized mutual correlation coefficient between the extracted watermark and the original watermark is calculated to obtain the watermark similarity. The normalized mutual correlation coefficient calculation formula is: NCC = Σ[(X-μx)(Y-μy)] / (σx·σy), where X and Y are the extracted watermark and the original watermark, respectively, μ is the mean, and σ is the standard deviation. The path integrity scoring model is constructed, and the path integrity scoring model takes into account the number of nodes, the distortion rate distribution and the path criticality. The path integrity score is obtained, and whether the power data asset has been tampered with is determined according to the path integrity score.

[0041] Specifically, a depth-first search algorithm is used to traverse the lineage graph, identify all the circulation path nodes that the power data assets pass through, store the node information in an ordered list, and record the dependencies between the nodes. According to the node type (such as substation, distribution line, user end), select appropriate watermark extraction parameters, including wavelet transform level and subband selection. For each circulation path node, a three-level discrete wavelet transform is used to perform multi-scale decomposition of the power data, and the embedded digital watermark information is extracted in the LL3 subband. The sliding window method is used for correlation detection, the window size is set to 1 / 4 of the watermark length, the step size is 1, and the watermark sequence is restored by calculating the local correlation peak. If the extraction fails, the BCH error correction code is used to repair the watermark, or a multiple watermark strategy is used to improve robustness. The watermark similarity is obtained by calculating the normalized mutual correlation coefficient between the extracted watermark and the original watermark. The formula for calculating the normalized cross-correlation coefficient is: NCC = Σ[(X-μx)(Y-μy]) / (σx·σy), where X and Y are the extracted watermark and the original watermark, respectively, μ is the mean, and σ is the standard deviation. The distortion rate is calculated according to the similarity value, and the distortion rate calculation formula is: distortion rate = 1-similarity. A comprehensive analysis is performed on the distortion rates of all circulation path nodes, and the weighted average method is used to calculate the overall distortion rate. The weight is determined according to the importance of the node in the path. A path integrity scoring model is constructed, and the path integrity score is obtained by considering the number of nodes, distortion rate distribution and path criticality. The distortion rate threshold and path integrity score threshold are set. If the calculated overall distortion rate exceeds the preset threshold or the path integrity score is lower than the threshold, it is determined that the power data asset has been tampered with, triggering a tampering alarm and recording the suspicious node information. If the distortion rate is lower than the threshold and the path integrity score is higher than the threshold, it is determined that the data asset integrity is not damaged.

[0042] Taking a regional power grid as an example, a depth-first search algorithm is used to traverse the lineage graph and identify a flow path containing 5 nodes: substation A, distribution line B, distribution transformer C, user meter box D and smart meter E. Watermark extraction parameters are selected according to the node type. Substations and distribution lines use 3-level wavelet transform and LL3 subband extraction; other nodes use 2-level wavelet transform and LL2 subband extraction. The voltage data of substation A is subjected to 3-level discrete wavelet transform, and 128-bit watermark information is extracted in the LL3 subband. A 64-bit sliding window with a step size of 1 is used to calculate the local correlation peak and restore the watermark sequence. If the extraction fails, the (127,64,10) BCH code is applied for error correction, which can correct up to 10 errors. The normalized mutual correlation coefficient between the extracted watermark and the original watermark is calculated, and the similarity is 0.92. The distortion rate is 1-0.92=0.08. The distortion rates of the five nodes were calculated respectively, and the results were [0.08, 0.10, 0.15, 0.12, 0.18]. The weighted average method was used to calculate the overall distortion rate, and the weights were set to [0.3, 0.25, 0.2, 0.15, 0.1], resulting in a weighted average distortion rate of 0.118. A path integrity scoring model was constructed, considering the number of nodes (5 points), distortion rate distribution (4 points) and path criticality (4 points), with a total score of 13 points. The distortion rate threshold was set to 0.15, and the path integrity score threshold was set to 12 points. Since 0.118<0.15 and 13>12, it was determined that the integrity of the power data asset was not damaged. Through this series of steps, the watermark verification and integrity judgment of the full circulation path of the power data asset were realized.

[0043] S106. When the watermark extracted from the power data node in the bloodline map is distorted, access control strategies of different granularities are designed to generate attribute encryption keys associated with user attributes, and fine-grained data authorization and permission management are performed. At the same time, a proxy re-encryption mechanism is introduced to convert the attribute encrypted ciphertext into a form that can be decrypted by different users, and through differential privacy, Laplace noise is injected during the ciphertext conversion process.

[0044] A distortion rate threshold is set according to the degree of watermark distortion of the power data node in the bloodline map, and the distortion rate threshold is used to divide the power data into three levels of high sensitivity, medium sensitivity and low sensitivity. The CP-ABE algorithm is used to generate associated attribute keys, and the attribute keys are generated according to attributes such as the user's position, department, and security level. The access control policy is mapped into an attribute tree, and the high-sensitivity data in the attribute tree is required to meet more attribute conditions. A proxy re-encryption mechanism is constructed, and the proxy re-encryption mechanism is implemented by generating a re-encryption key RK, where RK=H(SK_A, PK_B), SK_A is the original key, PK_B is the target user's public key, and H is a hash function. A differential privacy mechanism based on Laplace distribution is introduced in the re-encryption process, and the differential privacy mechanism injects noise into sensitive attributes, and the epsilon value of the noise is inversely proportional to the sensitivity of the data.

[0045] Specifically, according to the degree of watermark distortion of the power data node in the bloodline map, the distortion rate threshold is set: 0.1 for low distortion, 0.1-0.3 for medium distortion, and greater than 0.3 for high distortion. A multi-level access control strategy is designed based on the degree of distortion, and the data is divided into three levels of high sensitivity, medium sensitivity and low sensitivity, corresponding to different access rights and encryption strengths. The CP-ABE (Cipher Policy Attribute-Based Encryption) algorithm is used to generate associated attribute keys based on the user's position, department, security level and other attributes. The access control policy is mapped to an attribute tree, and high-sensitivity data is required to meet more attribute conditions. The attribute key is combined with the access policy tree to achieve fine-grained encryption and authorization of power data. A proxy re-encryption mechanism is constructed to generate a re-encryption key RK=H(SK_A, PK_B), where SK_A is the original key, PK_B is the target user's public key, and H is a hash function. The re-encryption key is used to convert the attribute-encrypted ciphertext into a form that can be decrypted by different users. A differential privacy mechanism based on Laplace distribution is introduced during the re-encryption process to inject noise into sensitive attributes. The epsilon value is set to be inversely proportional to the data sensitivity, with epsilon = 0.1 for high-sensitivity data, epsilon = 0.5 for medium-sensitivity data, and epsilon = 1 for low-sensitivity data. An attribute-based access control matrix is ​​designed, with rows representing users and columns representing data objects. The matrix element values ​​are {0, 1, 2}, representing no permission, read permission, and read-write permission, respectively. Combined with user attributes and data sensitivity, user permissions are dynamically adjusted to achieve fine-grained permission management. All access operations are recorded, including user ID, operation type, access object, timestamp, and result, to form a structured audit log.

[0046] Taking the distribution network data of a certain power company as an example, the voltage data of substation A is watermarked and the calculated distortion rate is 0.25, which belongs to the medium distortion level. According to the degree of distortion, the data is divided into medium sensitivity levels, and the access control policy is set to require users to have both "distribution department" and "engineer level". The CP-ABE algorithm is used to generate attribute keys, and the attribute set is {distribution department, engineer, substation A}. The access policy tree is constructed as ("distribution department" AND "engineer"), and the policy is embedded in the ciphertext. The voltage data is encrypted to obtain the ciphertext CT. A user Bob applies to access the data, and the system verifies that Bob's attribute set {distribution department, engineer, substation B} meets the access policy. Generate a re-encryption key RK = SHA256 (SK_A||PK_Bob), where SK_A is the original private key and PK_Bob is Bob's public key. Use RK to convert CT into a form that Bob can decrypt CT'. During the re-encryption process, Laplace noise is injected into the sensitive attribute "Substation A", and epsilon is set to 0.5. A 10x10 access control matrix is ​​constructed, with rows representing 10 users and columns representing 10 data objects. The permission value corresponding to Bob is set to 1 (read-only permission). The access log is recorded: {User ID: Bob_001 Operation: Read, Object: Substation A_Voltage Data, Time: 2023-04-01 10:30:15, Result: Success}. Through this series of steps, fine-grained access control, secure data sharing, and privacy protection based on watermark distortion are achieved.

[0047] S107. Based on the federated learning framework, data asset nodes in the bloodline graph are used as samples, and collaborative training is performed in a distributed environment. Modeling and analysis are performed on the data asset ciphertext after differential privacy is injected with Laplace noise, and secure multi-party computation of the data asset ciphertext is performed to conduct fusion analysis of power data assets in a confidential state.

[0048] The data asset node information in the bloodline map is obtained, and the data asset node is used as a participant in federated learning to build a distributed training architecture for horizontal federated learning. According to the sensitivity of the original data of the data asset node, the Laplace mechanism is used to inject noise into the sensitive features to obtain training samples with differential privacy protection. For the training samples with differential privacy protection, the Paillier homomorphic encryption algorithm is used for encryption processing to generate training samples in ciphertext form. Local model training is performed on the ciphertext domain using the training samples in ciphertext form to obtain encrypted gradient updates. If the encrypted gradient update is obtained, the Shamir secret sharing scheme is used to split the encrypted gradient update into n parts, and the threshold is set to t. The global gradient is calculated through the secure summation protocol, and the original gradient is reconstructed from any t parts of the encrypted gradient to obtain the plaintext global gradient.

[0049] Specifically, a distributed training architecture based on horizontal federated learning is constructed, and the data asset nodes in the bloodline graph are used as participants in federated learning. Each node retains the original data locally, and secure communication and model parameter exchange are carried out through the SSL / TLS protocol. Differential privacy processing is performed on the original data of each data asset node, and noise is injected into sensitive features through the Laplace mechanism. The noise size dynamically adjusts the privacy budget ε according to the data sensitivity. The high-sensitivity data ε=0.1, the medium-sensitivity ε=0.5, and the low-sensitivity ε=1 to generate training samples with privacy protection. The Paillier homomorphic encryption algorithm is used to encrypt the processed training samples, and a 2048-bit key length is selected to convert the samples into ciphertext form. Local model training is performed on the ciphertext domain to achieve specific tasks such as power load prediction, and encrypted gradient updates are obtained. The encrypted gradient is transmitted to the central aggregation server through a secure channel. The Shamir secret sharing scheme is used to implement secure multi-party computing, and the encrypted gradient is divided into n parts. The threshold is set to t, and any t parts can reconstruct the original gradient. The global gradient is calculated using a secure summation protocol, and the plaintext global gradient is obtained through a threshold decryption operation to update the global model parameters. The model is evaluated to calculate the prediction error and privacy protection effect, such as the trade-off between the epsilon value of differential privacy and model accuracy. If the preset indicators are met, the training is completed; otherwise, the differential privacy processing step is returned, and the parameters are adjusted and the iteration continues.

[0050] Taking a provincial power grid company as an example, a federated learning network consisting of 10 municipal power supply companies was constructed, and each node represented a data asset. The TLS 1.3 protocol was used for communication encryption, and the ECDHE algorithm was used for key exchange. Differential privacy processing was performed on the original power load data, and the ε of sensitive features such as user power consumption was set to 0.1, and the ε of secondary features such as weather data was set to 0.5. Paillier homomorphic encryption was used, with a key length of 2048 bits and an encryption function of E(m+)=(1+n)^m*r^n mod n^2, where n is the product of two large prime numbers and r is a random number. The LSTM model was trained in the ciphertext domain for power load forecasting. The input features included historical load, temperature, humidity, etc., and the output was the load forecast for the next 24 hours. After each round of training, the node transmitted the encrypted gradient to the central server through a secure channel. The (7,10) Shamir secret sharing scheme was used to divide the gradient into 10 parts, and any 7 parts could be reconstructed. Use the secure summation protocol Σi=1^10[gi], where [gi] is the encrypted gradient share of the i-th node. Obtain the plaintext global gradient through threshold decryption and update the model parameters. Evaluate the model performance and calculate the mean absolute percentage error MAPE. If MAPE<5% and the ε comprehensive value<1, the training is completed; otherwise, adjust the parameters and continue to iterate. Through this series of steps, the secure fusion analysis of power data assets and load forecasting under privacy protection are achieved.

[0051] It should be noted that the above examples are only some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A distributed power data asset management privacy protection encryption architecture design method, characterized in that: The method comprises: According to the statistical characteristics of power data assets, a content-based digital watermark algorithm is used to generate a robust and imperceptible watermark, and the watermark embedding process and extraction results are stored on the blockchain. Aiming at watermark embedding and extraction in distributed environment, a watermark protocol for secure multi-party computing is designed. By using secret sharing and blind watermark mechanism, privacy-preserving embedding and blind extraction of watermarks of power data assets are performed to obtain power data assets carrying watermarks. A graph-based data model is adopted, with watermarked power data assets as nodes and data flow relationships as edges, to construct a power data asset lineage graph; For the distributed storage of bloodline graphs, blockchain and distributed hash tables are used to build a P2P network topology, and the bloodline graph data is distributed and stored in each node of the alliance chain; Verify the digital watermark information embedded in all the circulation path nodes of the power data assets in the lineage map, calculate the similarity between the extracted watermark and the original watermark, calculate the distortion rate, and determine whether the power data assets have been tampered with; When the watermark extracted from the power data node in the bloodline graph is distorted, access control strategies of different granularities are designed to generate attribute encryption keys associated with user attributes, and fine-grained data authorization and permission management are performed. At the same time, a proxy re-encryption mechanism is introduced to convert attribute encrypted ciphertext into a form that can be decrypted by different users, and Laplace noise is injected into the ciphertext conversion process through differential privacy. Based on the federated learning framework, data asset nodes in the bloodline graph are used as samples, and collaborative training is carried out in a distributed environment. The data asset ciphertext after differential privacy injection of Laplace noise is modeled and analyzed, and secure multi-party computation of data asset ciphertext is performed to conduct fusion analysis of power data assets in a confidential state.

2. The method according to claim 1, wherein: According to the statistical characteristics of power data assets, a content-based digital watermark algorithm is used to generate a robust and imperceptible watermark, and the watermark embedding process and extraction results are stored on the blockchain, including: Performing time series analysis and spectrum analysis on the power data assets to obtain statistical feature vectors of the power data assets, wherein the statistical feature vectors include mean, variance, autocorrelation coefficient and power spectrum density; According to the statistical feature vector, a digital watermark algorithm is designed by using discrete cosine transform, and a watermark embedding parameter is obtained by adjusting the watermark embedding strength coefficient; Performing multi-scale decomposition on the power data asset using discrete wavelet transform to obtain low-frequency sub-bands; Selecting an embedding position in the low-frequency sub-band, encoding the watermark information using a spread spectrum watermarking technique, and obtaining encoded watermark information; Performing a hash operation on the watermark embedding parameter and the encoded watermark information to generate a unique identifier; Building a block data structure in combination with the timestamp information, and adding the block data structure to the blockchain network using a proof-of-stake consensus mechanism; Design a watermark extraction verification scheme based on the zk-SNARKs zero-knowledge proof protocol and deploy verification logic in blockchain smart contracts; If a watermark extraction request is received, executing the verification logic; The verification logic includes: obtaining the unique identifier from the blockchain network, performing a hash operation on the watermark information in the watermark extraction request, and obtaining a hash result; It is determined whether the hash result matches the unique identifier, and if so, it is determined that the watermark extraction request is valid.

3. The method according to claim 1, wherein: The watermarking protocol for secure multi-party computing is designed for watermark embedding and extraction in a distributed environment. The watermarking protocol uses secret sharing and blind watermarking mechanisms to perform privacy protection embedding and blind extraction of power data asset watermarks to obtain power data assets carrying watermarks, including: The Shamir threshold secret sharing scheme is used to split the watermark information, the threshold value is set to t, n watermark information fragments are obtained, and distributed to n distributed nodes; According to the watermark information fragmentation and the time series characteristics of the power data assets, the discrete cosine transform domain is selected as the watermark embedding space, and the transform coefficients are encrypted using the Paillier homomorphic encryption algorithm; Performing a watermark embedding operation in the encryption domain, selecting an intermediate frequency coefficient as an embedding position, adjusting the coefficient amplitude according to the watermark information slice, and using the additive property of homomorphic encryption to implement coefficient modification in the encryption domain; A watermark extraction protocol is designed through a blind watermark mechanism to extract watermark information in the encrypted domain, and the Schnorr zero-knowledge proof protocol is used to verify the existence of the watermark; If the verification is successful, the (t, n) threshold RSA signature scheme is used to verify the final result, generate an RSA key pair, obtain n private key shards, and distribute them to the n distributed nodes; Obtain partial signatures from t nodes among the n distributed nodes; Combining the t partial signatures to obtain a complete signature; Verify the validity of the complete signature.

4. The method according to claim 1, wherein: The graph-based data model is adopted, with the power data assets carrying watermarks as nodes and the data flow relationships as edges, to construct a power data asset lineage graph, including: Defining attribute fields of the nodes in the graph data model, wherein the attribute fields include asset identification, data type, timestamp, spatial location, and watermark information; Using Neo4j property graph database to store the node information, and using Cypher language to create the node structure; Constructing a data flow relationship edge according to the node information, and calculating a weight value of the relationship edge, wherein the weight value is determined by the data flow frequency, the data volume, and the data importance; Using a depth-first search algorithm to traverse the graph constructed by the relationship edges, identifying the blood relationship between data assets; The Louvain community detection algorithm is used to optimize the structure of the blood relationship graph, and redundant nodes and edges are merged to obtain an optimized blood relationship graph; Serializing the optimized blood relationship map into JSON format and storing it in the Neo4j database; Read the serialized bloodline graph from the Neo4j database and deserialize it into a graph structure; Using the D3.js visualization library to implement a graphical interface, the graph structure data is converted into a force-directed graph layout; For the nodes carrying watermarks in the force-directed graph layout, a watermark verification function is called to check the integrity of the watermark, and different colors are used in the graphical interface to mark nodes with normal and abnormal watermark states.

5. The method according to claim 1, wherein: The distributed storage of the bloodline map uses blockchain and distributed hash tables to build a P2P network topology, and stores the bloodline map data in each node of the alliance chain in a distributed manner, including: The Chord algorithm is used to construct a distributed hash table to serialize the bloodline map data into a key-value pair format, where the key is the node identifier and the value is the node attribute and adjacency relationship; According to the key-value pair format data, the mapping relationship between the data and the alliance chain node is calculated by a consistent hash function; Design a blockchain consensus mechanism based on a practical Byzantine fault-tolerant algorithm to synchronize the update operation of the bloodline graph data between alliance chain nodes; The consensus mechanism includes five stages: master node election, request broadcast, pre-preparation, preparation and submission, and the validity of data is verified through the voting mechanism; The bloodline graph data transmitted between nodes is encrypted using the RSA asymmetric encryption algorithm, and node identity authentication is achieved in combination with the zk-SNARK zero-knowledge proof protocol; In the zero-knowledge proof protocol, the prover generates a proof containing node identity information, and the verifier verifies the validity of the proof without obtaining specific identity information; Constructing a distributed index based on a graph structure, and using the topological features of the bloodline graph to establish an inverted index of the relationship between nodes; Design a distributed graph query algorithm to decompose complex bloodline graph query tasks into subqueries; In the distributed graph query algorithm, the query plan generator formulates the query path according to the topological structure characteristics of the graph; Subqueries are distributed to relevant alliance chain nodes through the P2P network, and query operations are executed in parallel; The MapReduce framework is used to summarize query results. The Map phase processes subquery results, and the Reduce phase merges the final query results.

6. The method according to claim 1, wherein: The verification of the digital watermark information embedded in all the circulation path nodes of the power data asset in the lineage map, by calculating the similarity between the extracted watermark and the original watermark, calculating the distortion rate, and judging whether the power data asset has been tampered with, includes: Using a depth-first search algorithm to traverse the lineage graph, identify all nodes along the circulation path through which the power data asset passes, and store the node information in an ordered list; Selecting watermark extraction parameters according to the node type, the watermark extraction parameters including wavelet transform series and subband selection; Performing multi-scale decomposition on the power data using a 3-level discrete wavelet transform, and extracting embedded digital watermark information in the LL3 subband; A sliding window method is used for correlation detection, where the window size of the sliding window method is set to 1 / 4 of the watermark length and the step size is 1; The watermark sequence is restored by calculating the local correlation peak value. If the extraction fails, the watermark is repaired using the BCH error correction code; Calculate the normalized cross-correlation coefficient between the extracted watermark and the original watermark to obtain the watermark similarity. The normalized cross-correlation coefficient calculation formula is: NCC = Σ[(X-μx)(Y- μy)] / (σx·σy), where X and Y are the extracted watermark and the original watermark, μ is the mean, and σ is the standard deviation; Constructing the path integrity scoring model, wherein the path integrity scoring model considers the number of nodes, the distortion rate distribution and the path criticality; The path integrity score is obtained, and whether the power data asset has been tampered with is determined according to the path integrity score.

7. The method according to claim 1, wherein: When the watermark extracted from the power data node in the bloodline map is distorted, access control strategies of different granularities are designed to generate attribute encryption keys associated with user attributes, and fine-grained data authorization and authority management are performed. At the same time, a proxy re-encryption mechanism is introduced to convert the attribute encrypted ciphertext into a form that can be decrypted by different users, and Laplace noise is injected into the ciphertext conversion process through differential privacy, including: A distortion rate threshold is set according to the degree of watermark distortion of the power data node in the lineage graph, and the distortion rate threshold is used to divide the power data into three levels of high sensitivity, medium sensitivity and low sensitivity; The CP-ABE algorithm is used to generate associated attribute keys, which are generated based on attributes such as the user's position, department, and security level; Mapping the access control policy into an attribute tree, wherein highly sensitive data in the attribute tree is required to satisfy more attribute conditions; Construct a proxy re-encryption mechanism, which is implemented by generating a re-encryption key RK, where RK = H (SK_A, PK_B), SK_A is the original key, PK_B is the target user's public key, and H is a hash function; A differential privacy mechanism based on Laplace distribution is introduced in the re-encryption process. The differential privacy mechanism injects noise into sensitive attributes, and the epsilon value of the noise is inversely proportional to the sensitivity of the data.

8. The method according to claim 1, wherein: Based on the federated learning framework, the data asset nodes in the bloodline graph are used as samples, collaborative training is performed in a distributed environment, modeling and analysis are performed on the data asset ciphertext after differential privacy is injected with Laplace noise, and secure multi-party computation of the data asset ciphertext is performed. The fusion analysis of power data assets is performed in a confidential state, including: Obtain the data asset node information in the lineage graph, and use the data asset node as a participant in federated learning to build a distributed training architecture for horizontal federated learning; According to the sensitivity of the original data of the data asset node, the Laplace mechanism is used to inject noise into the sensitive features to obtain training samples with differential privacy protection; For the training samples with differential privacy protection, the Paillier homomorphic encryption algorithm is used to perform encryption processing to generate training samples in ciphertext form; Performing local model training on the ciphertext domain using the ciphertext training samples to obtain encrypted gradient updates; If the encrypted gradient update is obtained, the encrypted gradient update is divided into n parts using the Shamir secret sharing scheme, and the threshold is set to t; The global gradient is calculated through a secure summation protocol, and the original gradient is reconstructed from any t parts of the encrypted gradient to obtain the plaintext global gradient.

Images