An AI-based enterprise digital marketing data security protection system

By constructing a semantic lineage graph of marketing data and introducing a privacy management interface, the problems of data loss and privacy compliance in existing data recovery systems have been solved, enabling accurate data repair and compliance with privacy protection requirements without interrupting business operations.

CN123116271APending Publication Date: 2026-11-17SHANDONG YUANXING NETWORK TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611271511.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing marketing data backup and recovery systems use a point in time as the smallest recovery granularity, which cannot identify the semantic relationships and pollution range between data records. This results in the loss of a large amount of legitimate data after a data poisoning attack, and it is difficult to keep the production system running without interruption and business operations uninterrupted during the recovery process. At the same time, it is difficult to balance the immutability of existing systems with privacy compliance.

Method used

A graph neural network is used to construct a semantic lineage graph of marketing data. Combined with an anomaly score iterative propagation mechanism, polluted data clusters are accurately identified, and targeted overwrite recovery is performed through a precise cleaning execution engine. At the same time, a privacy management interface and a streaming recovery channel are introduced to compare user deletion requests in real time to achieve the removal or desensitization of privacy data.

Benefits of technology

It enables precise repair of legitimate data after a data poisoning attack, reducing business impact, and meets privacy compliance requirements without compromising the immutability of backups, thus avoiding the problems of losing legitimate data and rewriting private data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116271A_ABST
    Figure CN123116271A_ABST
Patent Text Reader

Abstract

This invention discloses an artificial intelligence-based enterprise digital marketing data security protection system, relating to the fields of artificial intelligence and network data security technology. The system includes a production database, an incremental log collection module, a lineage graph construction module, an AI poisoning detection and tracing module, a WORM backup storage layer, and a precise cleaning execution engine. It constructs a data lineage graph through a graph neural network and iteratively propagates abnormal scores to accurately identify poisoned data clusters. It compiles fine-grained instructions to selectively overwrite the production database, avoiding the loss of legitimate data during overall rollback. The system also includes a privacy management interface, a privacy logic deletion index pool, and a streaming recovery channel. Under the premise that the backup data remains immutable, it compares and de-identifies the data recovered in the transmission state in real time, achieving refined data recovery and unified privacy compliance. It can be used for enterprise marketing data security protection and disaster recovery.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and network data security technology, specifically to an artificial intelligence-based enterprise digital marketing data security protection system. Background Technology

[0002] As enterprises deepen their digital transformation, marketing operations rely heavily on AI recommendation algorithms and user profiling systems. Related marketing data is characterized by its large volume, frequent updates, and the presence of a significant amount of user privacy information. To ensure data reliability, existing technologies generally employ a combination of full snapshot backups and incremental logs to back up production databases. In the event of data corruption, a specific snapshot and incremental log are selected based on a recovery point in time, and the entire database or a portion of the tables is rolled back to that point. Furthermore, to protect backup files from ransomware encryption, the industry commonly uses object storage's write-once-read-many mechanism, ensuring that backup files are physically unmodifiable and undeletable, guaranteeing the existence of historical copies available for recovery after an attack. In addition, graph neural networks have been applied to relational modeling and anomaly detection, enabling the propagation and aggregation of node anomaly information within the relational network based on the message passing mechanism of graph-structured data, demonstrating considerable engineering feasibility.

[0003] However, the aforementioned backup and recovery methods only use a point in time as the smallest recovery granularity, lacking the ability to identify semantic relationships and the scope of contamination between data records. Furthermore, there is a long-standing technical bias in the field, generally believing that the immutability of backup data and the user's legally requested right to be forgotten deletion are inherently mutually exclusive and incompatible in the same recovery pipeline. That is, if the backup is kept immutable, the deleted privacy data will inevitably be rewritten during recovery; if the backup is modified to satisfy the deletion request, the immutability mechanism required for ransomware protection will inevitably be destroyed. Therefore, existing technologies generally do not attempt to satisfy both simultaneously within the same system architecture. This invention overcomes this technical bias: if the backup is strictly kept immutable, the recovery operation may rewrite legally deleted privacy data into the production environment; if the backup file is modified to satisfy the deletion request, the immutability required for ransomware protection will be destroyed. Therefore, existing technologies cannot simultaneously achieve both refined data recovery and privacy compliance within the same system. To address this, we propose an artificial intelligence-based enterprise digital marketing data security protection system. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based enterprise digital marketing data security protection system.

[0005] The technical problem this invention aims to solve is that existing marketing data backup and recovery systems only use a point in time as the smallest recovery granularity and lack the ability to identify semantic relationships and pollution ranges between data records. After being subjected to covert and gradual data poisoning attacks, they can only indiscriminately discard and roll back all data after the recovery time point, resulting in the deletion of a large amount of legitimate marketing data generated during the poisoning period, causing high business costs. Furthermore, it is difficult to keep the production system running without interruption and business operations uninterrupted during the recovery process. At the same time, the write-once, read-many backup mechanism used by existing systems physically prohibits any modification to the backup files, while the recovery operation directly and indiscriminately replays the backup file content to the production database. As a result, when users request deletion of the right to be forgotten after the backup is generated, the private data that should have been deleted is rewritten into the production environment during the recovery process, creating compliance risks. If the backup files are modified to meet compliance requirements, the immutability of the backup files will be compromised, rendering the anti-ransomware mechanism ineffective. Under the existing technical architecture, it is difficult to satisfy both requirements simultaneously in the same system.

[0006] To address the aforementioned technical problems, this invention provides an artificial intelligence-based enterprise digital marketing data security protection system.

[0007] To address the first technical problem, the system provided by this invention includes a production database, an incremental log collection module, a lineage graph construction module, an AI poisoning detection and tracing module, a WORM backup storage layer, and a precise cleaning execution engine. The lineage graph construction module uses marketing data records as graph nodes and the generation, referencing, and update relationships between data records as graph edges. It employs a graph neural network to construct and continuously update the data lineage graph. When the AI ​​poisoning detection and tracing module detects a preset abnormal operating condition in the artificial intelligence marketing model, it calculates an initial abnormality score for the suspected abnormal graph nodes and utilizes... The message passing mechanism of a graph neural network is used to iteratively propagate the abnormal scores of each graph node along the edges of the data lineage graph to its neighboring graph nodes. The set of graph nodes whose abnormal scores reach a preset judgment threshold after iteration convergence is defined as the contaminated data cluster. The precise cleaning execution engine retrieves clean backup data generated before the contamination occurred from the WORM backup storage layer based on the primary key and field range of the data involved in the contaminated data cluster, extracts clean data slices of the corresponding fields, and compiles the recovery action into fine-grained SQL instructions for the contaminated data cluster. The contaminated records are then overwritten without interrupting the external service of the production database.

[0008] To address the second technical problem, the system provided by this invention further includes a privacy management interface, a privacy logic deletion index pool, and a streaming recovery channel. The privacy management interface is used to receive user requests to delete the right to be forgotten and to perform hash calculations on the corresponding data primary keys. The privacy logic deletion index pool is used to store the hash values ​​and the timestamps of the corresponding deletion requests. The streaming recovery channel, while ensuring that the WORM backup data remains immutable, bypasses and intercepts the recovery data stream in transit, and compares in real time whether the user of each record has applied for the right to be forgotten and whether the application time is earlier than the backup generation time of the record. If the condition is met, the record is removed or de-identified before being written to the production database.

[0009] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0010] 1. This invention constructs a semantic lineage graph of marketing data through a graph neural network and combines it with an anomaly score iterative propagation mechanism. After being subjected to a covert data poisoning attack, it can accurately identify the polluted data clusters, avoiding the loss of a large amount of legitimate marketing data generated during the poisoning period due to the overall snapshot rollback method. At the same time, the precise cleaning execution engine uses fine-grained SQL commands to perform targeted overwriting of the production database, so that the data recovery process can be completed without interrupting the external services of the production system, effectively reducing the impact of data recovery on business continuity.

[0011] 2. This invention maintains the immutability required for ransomware protection by keeping the WORM backup data unchanged, and introduces a privacy logic deletion index pool for real-time comparison and dynamic processing during the data recovery transmission process. This ensures that privacy data that should be deleted is removed or desensitized before being written to the production environment, thereby achieving privacy compliance of the recovery operation without compromising the immutability of the backup. This solves the technical problem of the mutual exclusion between the ransomware integrity requirements and the right to be forgotten compliance requirements in the same recovery pipeline. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall main process of the system in an embodiment of the present invention;

[0013] Figure 2 This is a schematic diagram of the AI-based poisoning detection, tracing, and cleaning process in an embodiment of the present invention;

[0014] Figure 3 This is a schematic diagram of the privacy compliance streaming recovery process in an embodiment of the present invention. Detailed Implementation

[0015] The following is in conjunction with the appendix Figure 1-3The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0016] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] In this embodiment, the AI-based enterprise digital marketing data security protection system provided by the present invention includes a production database, an incremental log collection module, a lineage graph construction module, an AI poisoning detection and tracing module, a WORM backup storage layer, a precision cleaning execution engine, a privacy management interface, a privacy logic deletion index pool, and a streaming recovery channel. The production database is used to store user profile data, order data, and behavior log data generated by the enterprise's digital marketing business, and provides business read and write services to the outside world. The incremental log collection module is connected to the production database, and parses the incremental change logs output by the production database in real time. It extracts the operation type, operation timestamp, operation account identifier, and related business fields such as user identifier, order identifier, and behavior feature vector from each data change record to form a structured data change event stream for use by the lineage graph construction module.

[0018] The lineage graph construction module is connected to the incremental log collection module. It uses user profile attribute nodes, order record nodes, and behavior log nodes as graph nodes, and the generation, reference, and update relationships between data records as graph edges. It uses a graph neural network to perform vectorized encoding learning on each graph node to construct a dynamic data lineage graph covering all marketing data. The graph is periodically synchronized and stored in the graph database for subsequent real-time query and update calls.

[0019] The AI ​​poisoning detection and tracing module is connected to the lineage graph construction module and the model effect monitoring unit to continuously monitor the key performance indicators and input data distribution of the artificial intelligence marketing model (such as recommendation model and user profile model). When the performance indicator decline exceeds the preset threshold, or the input data distribution drifts more than the preset threshold compared to the historical baseline, it is determined that the artificial intelligence marketing model has a preset abnormal operation and triggers data poisoning tracing processing.

[0020] After triggering the source tracing process, the AI ​​poisoning detection and source tracing module calculates an initial anomaly score for suspected abnormal graph nodes in the data lineage graph. The calculation method is as follows: obtain the actual feature vector of the data record corresponding to the graph node and the predicted reconstruction value of the graph node's features by the baseline model trained based on historical normal data, calculate the Euclidean distance between the two, and normalize the Euclidean distance using the standard deviation of the historical reconstruction error corresponding to the data type to which the graph node belongs. The value obtained after normalization is used as the initial anomaly score of the graph node. The above calculation method can be expressed as the following formula:

[0021] ;

[0022] in, Represents graph nodes In iteration number The initial abnormal score of the round, Represents graph nodes The actual feature vector corresponding to the data record, This indicates the baseline model trained on historical normal data for graph nodes. Predicted reconstructed values ​​of features Represents graph nodes The standard deviation of the historical reconstruction error corresponding to the data type is used to normalize outlier metrics. The Euclidean norm of a vector;

[0023] After obtaining the initial anomaly scores for each suspected abnormal graph node, the AI ​​poisoning detection and tracing module utilizes the message passing mechanism of graph neural networks to iteratively propagate the anomaly scores of each graph node along the edges of the data lineage graph to its neighboring graph nodes. The anomaly score of each graph node after this iteration is updated as follows: the initial anomaly score of the graph node is weighted and aggregated with the anomaly scores of its neighboring graph nodes in the previous iteration according to the weights of the lineage-related edges. This aggregated value is then weighted and combined according to a preset propagation attenuation coefficient to obtain the anomaly score of the graph node after this iteration. The above iterative propagation method can be expressed as the following formula:

[0024] ;

[0025] in, Represents graph nodes In the Abnormal scores during round iterations The graph nodes obtained from the aforementioned calculations The initial abnormal score, Represents graph nodes The set of neighbor graph nodes in a data lineage graph Represents graph nodes Its neighboring graph nodes The weight of the blood relation edges between them; the larger the weight value, the stronger the relationship between the two. The propagation attenuation coefficient is a value greater than 100%. and less than This is used to control the degree of influence of abnormal information of neighboring graph nodes on the abnormal score of this graph node. The above iterative process continues until the change in the abnormal score of each graph node in two adjacent iterations is less than the convergence threshold preset according to the historical iteration convergence data. The convergence threshold is used to determine whether the abnormal score propagation calculation has reached a stable state.

[0026] After iterative convergence, the AI ​​poisoning detection and tracing module sets an anomaly score judgment threshold, and judges the set of graph nodes whose anomaly scores reach the judgment threshold as suspected contaminated data clusters. Furthermore, combined with the time window constraint of the data poisoning behavior, the suspected contaminated data clusters are filtered in the time dimension to remove graph nodes that do not meet the poisoning injection characteristics in time, thus obtaining the final contaminated data clusters.

[0027] The precise cleaning execution engine is connected to the AI ​​poisoning detection and tracing module and the WORM backup storage layer. After receiving the judgment result of the contaminated data cluster, it retrieves the complete backup version generated before the contamination occurred and not marked as contaminated in the WORM backup storage layer according to the data primary key and field range involved in the contaminated data cluster. From this, it extracts clean data slices of the corresponding fields. The AI ​​poisoning detection and tracing module is also used to synchronously write the identification result into the contamination status mark field of the corresponding backup version in the WORM backup storage layer after identifying the contaminated data cluster. The precise cleaning execution engine further automatically compiles the recovery action into a surface. Fine-grained SQL update or delete commands are sent to contaminated data clusters to perform targeted overwriting and repair on the located contaminated records without interrupting the production database's external services. Corresponding cleaning operation logs are generated for subsequent auditing. After cleaning is completed, the precision cleaning execution engine re-acquires the AI ​​marketing model evaluation indicators corresponding to the region where the affected data clusters are located to verify whether the data poisoning effect has been eliminated. If the evaluation indicators have not recovered to the preset normal range, the AI ​​poisoning detection and tracing module is triggered to expand the scope of anomaly investigation and re-execute the anomaly score calculation and contaminated data cluster delineation processing until the repair effect meets the preset requirements.

[0028] The WORM backup storage layer connects to the production database, the precision cleaning execution engine, and the streaming recovery channel. It uses the write-once-read-many mechanism of object storage to physically save full backup data and incremental backup data. At the physical level, modification and deletion operations on the written backup data are prohibited, thus ensuring that even if the production environment is attacked by ransomware, there are still tamper-proof historical backup versions available for recovery.

[0029] The privacy management interface is used to receive user requests to delete the right to be forgotten. After receiving the deletion request, it performs a hash calculation on the data primary key (such as user identifier) ​​associated with the user to obtain the corresponding hash value. The privacy logic deletion index pool is connected to the privacy management interface and stores the above hash value and the timestamp of the corresponding deletion request in the form of an in-memory hash table for comparison in subsequent recovery processes.

[0030] The streaming recovery channel connects to the WORM backup storage layer, the privacy logic deletion index pool, and the production database. It has a built-in de-identification operator. When partial, precise recovery is needed due to data poisoning, or a full recovery is required for other reasons, the streaming recovery channel reads the corresponding range of backup data from the WORM backup storage layer, forming a recovery data stream in transit, ready to be written to the production database. The de-identification operator extracts the user identifier field of each record in the recovery data stream in real time and calculates the corresponding hash value. This hash value is then compared one by one with the hash values ​​stored in the privacy logic deletion index pool. The above comparison and decision logic can be expressed as the following two conditional expressions:

[0031] ;

[0032] ;

[0033] in, This indicates that a pending record in the data stream is being restored. Indicates the record The hash value calculated from the user identifier field. This indicates that the privacy logic has deleted the set of user identifier hash values ​​that have been processed from the index pool. Representing records The timestamp corresponding to the generation of the backup data This indicates the timestamp of the user's right to be forgotten deletion request recorded in the privacy logic deletion index pool. Indicates the record The processing result is as follows: when the hash value hits the privacy logic deletion index pool and the generation timestamp of the corresponding backup data is earlier than the deletion request acceptance timestamp, it means that the record should have been deleted at the time of recovery. The de-identification operator performs one of the following processing on the record: removal processing or de-identification processing, and then writes it into the production database. Removal processing means not writing the record into the production database, and de-identification processing means performing one of the following transformations on the privacy-sensitive fields in the record: mask replacement or irreversible hash transformation, and then writes it into the production database. Otherwise, the de-identification operator writes the record directly into the production database.

[0034] The aforementioned removal and desensitization processes only apply to recovery data in memory that is in transit. They do not modify the backup data already written to the WORM backup storage layer or its hash chain. Thus, privacy compliance of the recovery operation is achieved without compromising the immutability of the backup data. During the recovery process, the streaming recovery channel continuously generates desensitization processing logs, recording the user identifier, processing method, and processing time of the processed records. After all the recovery data streams are processed and written to the production database, the system outputs a recovery report, which includes statistics such as the total number of recovered records, the number of records that underwent desensitization and removal processing, and index pool hit status, for manual review and confirmation of recovery compliance.

[0035] In an alternative embodiment of the present invention, the data poisoning tracing process may not use graph neural networks for lineage mapping and anomaly propagation calculation. Instead, a rule engine and statistical models (such as sliding window statistical offset detection and KL divergence distribution drift detection) are used to determine data anomalies. A lineage relationship table that records foreign keys and references between data records in a relational database is used instead of a graph database to store lineage relationships. The correlation diffusion judgment of the anomaly range is achieved through relation queries. This alternative embodiment can still achieve approximate location and accurate cleaning of contaminated data in scenarios with limited computing resources or relatively simple data association structures.

[0036] In another alternative embodiment of the present invention, the privacy compliance recovery process may not adopt the real-time interception method based on the streaming computing framework. Instead, before the formal recovery operation is performed, the backup data set to be recovered is first scanned offline in batches. Records that hit the privacy logic deletion index pool and meet the time sequence determination conditions are pre-generated into a list to be de-identified. Then, the recovery copy of the backup data is processed in batches for de-identification or pre-processing for removal based on the list before being written into the production database. This alternative embodiment is suitable for application scenarios with relatively relaxed recovery time requirements and allow for additional pre-processing time windows.

[0037] Through the above system architecture and processing mechanism, this invention, after being subjected to a covert data poisoning attack, can achieve fine-grained positioning and precise repair of the contaminated data without interrupting the external services of the production database. This avoids the loss of legitimate marketing data generated during the poisoning period due to overall rollback. At the same time, while maintaining the immutability of backup data, it achieves a compliant response to the user's right to be forgotten during the recovery operation, solving the technical problem of mutual exclusion between anti-ransomware integrity requirements and privacy compliance requirements in the same recovery pipeline.

[0038] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An artificial intelligence-based enterprise digital marketing data security protection system, characterized in that, include: A production database stores enterprise digital marketing data and provides business read / write services. An incremental log collection module, connected to the production database, parses the incremental change logs in real time, extracting operation types, timestamps, account identifiers, and associated business fields from the data change records to form a structured data change event stream. A lineage graph construction module, connected to the incremental log collection module, uses marketing data records as graph nodes and the generation, referencing, and update relationships between data records as edges. It employs a graph neural network to vectorize and encode each graph node, constructing and continuously updating a data lineage graph. An AI poisoning detection and tracing module, connected to the lineage graph construction module, calculates initial anomaly scores for suspected abnormal graph nodes in the data lineage graph when a preset abnormal operation is detected in the AI ​​marketing model. It then uses the message passing mechanism of the graph neural network to transmit the scores of each graph node. Anomaly scores are iteratively propagated along the edges of the data lineage graph to their neighboring graph nodes. The set of graph nodes whose anomaly scores reach a preset anomaly score threshold after iterative convergence is defined as a contaminated data cluster. The WORM backup storage layer, connected to the production database, is used to physically save full backup data and incremental backup data in a write-once-read-many manner. The precision cleaning execution engine, connected to the AI ​​poisoning detection and tracing module and the WORM backup storage layer, is used to retrieve, based on the primary key and field range of the data involved in the contaminated data cluster, a complete backup version generated before the contamination of the corresponding contaminated data cluster and not marked as contaminated, from the WORM backup storage layer. Clean data slices of the corresponding fields are extracted, and the recovery action is compiled into a fine-grained instruction selected from SQL update instructions and SQL delete instructions. Without interrupting the external services of the production database, targeted overwriting is performed on the records corresponding to the contaminated data cluster.

2. The enterprise digital marketing data security protection system based on artificial intelligence according to claim 1, characterized in that, The AI ​​poisoning detection and tracing module calculates the initial anomaly score of the suspected abnormal graph node as follows: it obtains the Euclidean distance between the actual feature vector of the data record corresponding to the graph node and the predicted reconstruction value of the graph node feature by the baseline model trained based on historical normal data, and normalizes the Euclidean distance by using the standard deviation of the historical reconstruction error corresponding to the data type to which the graph node belongs, and uses the value obtained after normalization as the initial anomaly score.

3. The enterprise digital marketing data security protection system based on artificial intelligence according to claim 2, characterized in that: When the AI ​​poisoning detection and tracing module iteratively propagates the abnormal scores of each graph node to its neighboring graph nodes along the edges of the data lineage graph, it determines the abnormal score of each graph node after the current iteration in the following manner: the initial abnormal score of the graph node and the abnormal scores of each neighboring graph node in the previous iteration are weighted and aggregated according to the lineage association edge weights, and then weighted and combined according to a preset propagation attenuation coefficient to obtain the abnormal score of the graph node after the current iteration. The propagation attenuation coefficient is greater than 0 and less than 1.

4. The enterprise digital marketing data security protection system based on artificial intelligence according to claim 1, characterized in that: After identifying the contaminated data cluster, the AI ​​poisoning detection and tracing module is also used to perform time-dimensional filtering on the contaminated data cluster based on the time window constraint of the data poisoning behavior, and to remove graph nodes that do not meet the poisoning injection characteristics in terms of time.

5. The enterprise digital marketing data security protection system based on artificial intelligence according to claim 1, characterized in that: It also includes a privacy management interface, a privacy logic deletion index pool, and a streaming recovery channel. The privacy management interface is used to receive user requests to delete the right to be forgotten and to perform a hash calculation on the primary key associated with the user's data. The privacy logic deletion index pool is connected to the privacy management interface and is used to store the hash value obtained from the hash calculation and the timestamp of the corresponding deletion request. The streaming recovery channel is connected to the WORM backup storage layer, the privacy logic deletion index pool, and the production database. It has a built-in desensitization operator, which is used to read the recovery data stream formed in the transmission state and ready to be written to the production database from the WORM backup storage layer when performing a recovery operation. For each record in the recovery data stream, the user identifier field to which the record belongs is extracted in real time and the corresponding hash value is calculated. The hash value is compared with the hash value stored in the privacy logic deletion index pool. If the comparison matches and the generation timestamp of the backup data corresponding to the record is earlier than the timestamp of the record in the privacy logic deletion index pool, the record is processed by one of the removal processing and desensitization processing and then written to the production database. If the condition is not met, the record is written to the production database.

6. The enterprise digital marketing data security protection system based on artificial intelligence according to claim 1, characterized in that: The precision cleaning execution engine is also used to obtain the AI ​​marketing model evaluation index corresponding to the area where the contaminated data cluster is located after the targeted overwrite is completed. If the evaluation index does not recover to the normal range determined by the baseline model trained based on the historical normal data, the AI ​​poisoning detection and tracing module is triggered to expand the scope of anomaly investigation and re-execute the anomaly score calculation and contaminated data cluster delineation processing.

7. An artificial intelligence-based enterprise digital marketing data security protection system, characterized in that, include: A production database stores enterprise digital marketing data and provides business read / write services. A WORM backup storage layer, connected to the production database, physically saves full and incremental backup data using a write-once-read-many approach, physically prohibiting modification and deletion operations on already written backup data. A privacy management interface receives user requests to delete the right to be forgotten and performs hash calculations on the primary key associated with that user. A privacy logic deletion index pool, connected to the privacy management interface, stores the hash value obtained from the hash calculation and the timestamp of the corresponding deletion request. A streaming recovery channel connects to the WORM backup storage layer, the privacy logic deletion index pool, and the production database, and includes a built-in de-identification operator to retrieve data from the production database during recovery operations. The WORM backup storage layer reads the recovery data stream that is in transit and ready to be written to the production database. For each record in the recovery data stream, the user identifier field to which the record belongs is extracted in real time and the corresponding hash value is calculated. The hash value is compared with the hash value stored in the privacy logic deletion index pool. If the comparison matches and the generation timestamp of the backup data corresponding to the record is earlier than the acceptance timestamp of the record in the privacy logic deletion index pool, the record is processed by one of the removal processing and de-identification processing and then written to the production database. If the condition is not met, the record is written to the production database. The removal processing and the de-identification processing only apply to the recovery data stream in transit and do not modify the backup data already written in the WORM backup storage layer.

8. The enterprise digital marketing data security protection system based on artificial intelligence according to claim 7, characterized in that: The desensitization process includes one of the following transformation processes: masking and irreversible hashing of privacy-sensitive fields in the record.

9. The enterprise digital marketing data security protection system based on artificial intelligence according to claim 7, characterized in that: The streaming recovery channel is also used to generate a desensitization processing log after the recovery process is completed. The desensitization processing log records the user identifier, processing method, and processing time of the processed record.