A data encryption privacy protection storage system for an AI business social platform

Abstract

The application relates to the technical field of data caching, in particular to a data encryption privacy protection storage system for an AI business social platform, which comprises a storage sample selection module, a clustering combination analysis module, an optional node determination module and a storage execution module. The clustering combination analysis module is used to determine whether to perform fragmentation processing based on sensitive content concentration according to a storage interaction sample category, and to determine a clustering combination based on a graph difference degree and a user portrait difference degree. The optional node determination module is used to determine an optional node based on whether to be in a stop time window. The storage execution module is used to allocate the optional node to each clustering combination according to a sensitive reference value and an associated interaction coefficient. The storage optimization module is used to determine whether to be in a sensitive accumulation state according to a risk backlog coefficient and a search hotspot connection degree, and to perform sensitive combination delay storage when being in the sensitive accumulation state. The application can reduce the risk of privacy leakage.

Description

Technical Field

[0001] This invention relates to the field of data caching technology, and in particular to a storage system for data encryption and privacy protection in an AI business social platform. Background Technology

[0002] With the rapid development of artificial intelligence technology, AI-driven social platforms are profoundly changing business interaction models. However, interactive AI platforms suffer from problems such as insufficient security control of storage nodes and risks of privacy data leakage, resulting in poor data storage effectiveness and risk adaptive handling capabilities. Therefore, improving data storage security is a problem that urgently needs to be solved by those skilled in the art.

[0003] Chinese Patent Publication No. CN110602147A discloses a data encryption and security storage method, system, and storage medium based on a cloud platform. The method includes: original data segmentation, establishment of a key distribution system and server storage cluster, calculation of storage node weights, storage of original data files and unit files, and data reconstruction when the primary storage node or replica storage node is abnormal, and a matching storage system is configured. It is evident that the above technical solution has the following problems: it does not consider the privacy leakage risk between multiple data samples, and it cannot adaptively adjust the storage strategy according to the actual application scenario, resulting in poor security and risk adaptability of sensitive data storage. Summary of the Invention

[0004] To address this, the present invention provides a storage system for data encryption and privacy protection in AI business social platforms, which overcomes the problems in existing technologies that do not consider the risk of privacy leakage between multiple data samples, cannot adapt storage strategies to actual application scenarios, and result in poor security and risk adaptability of sensitive data storage.

[0005] To achieve the above objectives, the present invention provides a storage system for data encryption and privacy protection in an AI business social platform, comprising: The storage sample selection module is used to determine the storage interaction samples based on the knowledge gain coefficient and the question depth value; The clustering and combination analysis module is used to determine the category of stored interaction samples based on intent jump coefficient and industry sensitivity, and to determine whether to perform segmentation based on the concentration of sensitive content based on the category of stored interaction samples, and to determine clustering combination based on graph difference and user profile difference. The optional node determination module is used to determine whether to disable a node based on the desensitization time inflation coefficient and the sensitivity of the interaction, and to determine whether to reduce the duration of the disabling time window based on the sensitivity dilution and the intermittent entropy of the disabling request, and to determine the optional node based on whether it is in the disabling time window. The storage execution module is used to allocate optional nodes for each cluster combination based on the sensitivity reference value and the correlation interaction coefficient; The storage optimization module is used to determine whether the system is in a sensitive accumulation state based on the risk backlog coefficient and the search hotspot connectivity, and to perform sensitive combination delayed storage when the system is in a sensitive accumulation state.

[0006] Furthermore, the storage sample selection module determines that interactive samples with a knowledge gain coefficient greater than a preset knowledge gain coefficient and a question depth value greater than a preset question depth value are stored interactive samples.

[0007] Furthermore, the clustering and combination analysis module determines that stored interaction samples with an intent jump coefficient greater than or equal to a preset intent jump coefficient or a peer sensitivity greater than or equal to a preset peer sensitivity are highly sensitive samples. The clustering and combination analysis module determines that stored interaction samples with an intent jump coefficient less than a preset intent jump coefficient and a peer sensitivity less than a preset peer sensitivity are low-sensitivity samples.

[0008] Furthermore, the clustering and combination analysis module determines whether to perform segmentation processing based on the concentration of sensitive content for highly sensitive samples.

[0009] Furthermore, the optional node determination module determines to disable a node if the desensitization time inflation coefficient is greater than the preset desensitization time inflation coefficient or the sensitivity interaction degree is greater than the preset sensitivity interaction degree, and sets the disabling time window duration to the base duration. The optional node determination module determines that nodes whose desensitization time dilation coefficient is less than or equal to the preset desensitization time dilation coefficient and whose sensitivity interaction degree is less than or equal to the preset sensitivity interaction degree do not need to be deactivated.

[0010] Furthermore, the optional node determination module determines to reduce the duration of the shutdown time window if the sensitive dilution is greater than the preset sensitive dilution and the intermittent entropy of the shutdown request is less than the preset intermittent entropy of the shutdown request.

[0011] Furthermore, the optional node determination module determines nodes that are not in the shutdown time window as optional nodes.

[0012] Furthermore, the storage execution module allocates optional nodes for each cluster combination based on the sensitive reference value and the correlation interaction coefficient; The storage execution module determines the allocation priority coefficient of each cluster combination based on the sensitive reference value, and determines the optional nodes of each cluster combination based on the association interaction coefficient. Among them, the allocation priority coefficient of a single cluster combination is positively correlated with the sensitivity reference value, and the optional nodes allocated to a single cluster combination are the optional nodes with the smallest association interaction coefficient with that cluster combination.

[0013] Furthermore, if the storage optimization module responds to a risk backlog coefficient greater than a preset risk backlog coefficient or a search hotspot connectivity greater than a preset search hotspot connectivity, it determines that it is in a sensitive backlog state.

[0014] Furthermore, the storage optimization module performs sensitive combined delayed storage; The storage optimization module performs delayed storage for each sensitive combination, and the delay time corresponding to each sensitive combination is positively correlated with the risk assessment value.

[0015] Compared with the prior art, the beneficial effects of the present invention are that, in the technical solution of the present invention, high-value interaction samples are effectively screened by knowledge gain coefficient and question depth value, and storage and desensitization resources are focused on complex interactions that contribute greatly to the expansion of knowledge graph and have long reasoning chains, avoiding the waste of resources caused by indiscriminate processing of all original question and answer data, and achieving a precise match between privacy protection investment and data value.

[0016] Furthermore, this invention categorizes interactive samples into high-sensitivity and low-sensitivity categories based on intent jump coefficients and industry sensitivity. High-sensitivity samples are then finely segmented based on the concentration of sensitive content, enabling fine-grained isolation of sensitive data with frequent topic shifts or high overlap in industry knowledge. Simultaneously, clustering combinations are constructed based on graph differences and user profile differences to ensure that the knowledge structure and user characteristics of data segments within the same combination are significantly different, effectively suppressing the risk of cross-leakage and amplification of sensitive information due to excessive aggregation or homogenization.

[0017] Furthermore, this invention dynamically determines whether a node should be disabled and the duration of the disabling time window by using the desensitization time dilation coefficient and the sensitivity interaction degree. It also adaptively shortens the disabling window based on the sensitivity dilution degree and the intermittent entropy of the disabling request, thereby achieving dynamic control of node availability. This allows nodes with uneven processing capacity or excessive coupling with the data to be processed to exit the service in a timely manner. At the same time, it determines the allocation priority coefficient of cluster combination by using the sensitivity reference value and selects the most suitable optional node by combining the association interaction coefficient, thus achieving dual optimization of storage load and sensitivity isolation requirements.

[0018] Furthermore, this invention comprehensively determines whether a system is in a sensitive accumulation state by combining the risk backlog coefficient with the correlation of public opinion hotspots. In the accumulation state, cluster combinations with high sensitivity reference values ​​are stored with a delay, and the delay time is positively correlated with the risk assessment value. This is beneficial for judging the overall risk situation from two dimensions: system load pressure and the coupling of public opinion hotspots, avoiding misjudgments or omissions caused by a single indicator. At the same time, in the sensitive accumulation state, the processing pace of high-density sensitive data is actively slowed down to buy buffer time for the system and significantly reduce the probability of sudden privacy leaks.

[0019] Furthermore, this invention adjusts the deactivation time window by using the intermittent entropy of deactivation requests and sensitivity dilution, so that the node deactivation time depends not only on its own performance and interaction sensitivity, but also on the sparseness of the data segment distribution in the system and the temporal regularity of deactivation requests. This achieves collaborative optimization from isolated node evaluation to global system stability, effectively avoiding information cross-leakage caused by implicit relationships between tasks, and significantly improving the data encryption and privacy protection capabilities in complex social platform scenarios. Attached Figure Description

[0020] Figure 1 This is a module connection diagram of a storage system for data encryption and privacy protection in an AI business social platform, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating how the storage interaction sample category is determined based on the intent jump coefficient and industry sensitivity, according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating how an embodiment of the present invention determines whether to reduce the duration of the downtime window based on the sensitivity dilution and the intermittent entropy of the downtime request. Figure 4 This is a flowchart illustrating how an embodiment of the present invention determines whether a region is in a sensitive accumulation state based on the risk accumulation coefficient and the search hotspot connectivity. Detailed Implementation

[0021] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] Please see Figure 1 The diagram shown is a module connection diagram of a storage system for data encryption and privacy protection in an AI business social platform according to an embodiment of the present invention. This embodiment of the present invention provides a storage system for data encryption and privacy protection in an AI business social platform, comprising: The storage sample selection module is used to determine the storage interaction samples based on the knowledge gain coefficient and the question depth value; The clustering and combination analysis module is used to determine the category of stored interaction samples based on intent jump coefficient and industry sensitivity, and to determine whether to perform segmentation based on the concentration of sensitive content based on the category of stored interaction samples, and to determine clustering combination based on graph difference and user profile difference. The optional node determination module is used to determine whether to disable a node based on the desensitization time inflation coefficient and the sensitivity of the interaction, and to determine whether to reduce the duration of the disabling time window based on the sensitivity dilution and the intermittent entropy of the disabling request, and to determine the optional node based on whether it is in the disabling time window. The storage execution module is used to allocate optional nodes for each cluster combination based on the sensitivity reference value and the correlation interaction coefficient; The storage optimization module is used to determine whether the system is in a sensitive accumulation state based on the risk backlog coefficient and the search hotspot connectivity, and to perform sensitive combination delayed storage when the system is in a sensitive accumulation state.

[0023] In this embodiment of the invention, the interaction samples are taken from the multi-round question and answer text of the AI ​​business social platform in the current storage period. A single interaction sample includes the content of multiple questions asked by a single user in the current storage period and the corresponding response text generated by the platform.

[0024] In this embodiment of the invention, the AI ​​business social platform is an intelligent business dialogue system built based on an initial knowledge graph, external data sources, and an interactive response generation model. The external data sources include, but are not limited to, external web pages, public databases, industry API interfaces, and real-time news feeds, used to acquire real-time business information, dynamic market data, and publicly available industry knowledge not covered by the initial knowledge graph. It is important to note that copyright and data authorization verification must be completed before accessing external data sources, and the scraping of unauthorized commercial data is prohibited.

[0025] In this embodiment of the invention, the storage period is the basic time unit for the system to perform clustering and risk analysis on the interaction samples. Within each storage period, the system performs a determination on the stored interaction samples of the collected interaction samples. The duration of a single storage period is set to a preset duration. As an example and not a limitation, the preset duration is set to 1 hour in this embodiment of the invention.

[0026] In this embodiment of the invention, several nodes are included, which are processing units with independent computing and storage resources. These nodes are adjacent to the selectable node determination module, the storage execution module, and the clustering and combination analysis module, respectively, and are used to de-identify the data fragments and store them. When de-identifying the data fragments, privacy information in the data fragments is removed or obfuscated. The specific de-identification process is not described in detail here, as it is a common technique in the art.

[0027] In this embodiment of the invention, the connection layout between each functional module and the storage node is as follows: The clustering and combinatorial analysis module is connected to several nodes and the sample selection storage module. An optional node determination module is connected to the plurality of nodes; The storage execution module is connected to the plurality of nodes, the optional node determination module, and the clustering combination analysis module, respectively. The storage optimization module is connected to both the clustering combination analysis module and the storage execution module.

[0028] In this embodiment of the invention, at a preset time, each node updates the data fragments that have completed the de-identification process for the initial knowledge graph and deletes them from local storage; wherein, the preset time can be configured by the user according to the system operation requirements; the knowledge graph adopts an incremental update mechanism, only writing new entities and entity relationships into the graph, without overwriting the original historical graph data.

[0029] In this embodiment of the invention, the initial knowledge graph is obtained as follows: First, a preset number of structured case corpora from experts are acquired. These structured case corpora include, but are not limited to, case background, core dilemmas, decision-making processes, and feedback on the final solution results. Entity extraction is performed using a domain-adaptive pre-trained language model. Then, a remote supervised relation extraction model combined with an attention mechanism is used to extract the relationships between entities. The extracted entities are used as nodes of the knowledge graph, and the extracted relationships are used as edges of the knowledge graph to construct the initial knowledge graph. The domain-adaptive pre-trained language model includes, but is not limited to, BioBERT and PubMedBERT. The remote supervised relation extraction model includes, but is not limited to, PCNN and BERT-RE. The construction of the knowledge graph is a common technique used by those skilled in the art, and will not be elaborated further.

[0030] Specifically, the storage sample selection module determines that interactive samples with a knowledge gain coefficient greater than a preset knowledge gain coefficient and a question depth value greater than a preset question depth value are stored interactive samples.

[0031] In this embodiment of the invention, the knowledge gain coefficient is determined based on the ratio of the number of newly added nodes in a single interaction sample to the total number of nodes in the interaction sample. The knowledge gain coefficient reflects the degree to which the interaction sample contributes to the information increment of the system knowledge graph.

[0032] For a single interaction sample, the newly added node is an entity node that did not appear in the initial knowledge graph in the entity extraction result performed by calling the domain adaptive pre-trained language model. The total number of nodes is the total number of entity nodes extracted in this entity extraction.

[0033] Question depth, which is the number of question-and-answer text rounds contained in a single interaction sample, reflects the degree to which the user explores the question in depth and the complexity of the interaction during the process.

[0034] In this embodiment of the invention, the larger the knowledge gain coefficient and question depth value of the interaction sample, the higher and longer the knowledge contribution rate and reasoning chain of the interaction sample. Therefore, the greater the demand for low-cost and rapid expansion of knowledge boundaries, the smaller the preset knowledge gain coefficient and preset question depth value should be. Preferably, historical working conditions that meet the usage requirements are obtained and stored. The average value of the knowledge gain coefficient corresponding to each stored interaction sample in the historical working conditions is 0.25, and the average value of the question depth value is 4, which are respectively denoted as the preset knowledge gain coefficient and preset question depth value.

[0035] Wherein, the historical working condition in which the storage meets the usage requirements is the storage period in which the number of cluster combinations without desensitization is less than the preset number and no data leakage event has occurred. As an example and not a limitation, in this embodiment of the invention, the preset number is set to 10% of the total number of cluster combinations in the corresponding storage period.

[0036] Please see Figure 2 As shown, it is a flowchart of an embodiment of the present invention for determining the category of stored interaction samples based on the intent jump coefficient and industry sensitivity.

[0037] Specifically, the clustering and combination analysis module determines that stored interaction samples with an intent jump coefficient greater than or equal to a preset intent jump coefficient or a peer sensitivity greater than or equal to a preset peer sensitivity are highly sensitive samples. The clustering and combination analysis module determines that stored interaction samples with an intent jump coefficient less than a preset intent jump coefficient and a peer sensitivity less than a preset peer sensitivity are low-sensitivity samples.

[0038] In this embodiment of the invention, the intent jump coefficient is determined based on the sum of the semantic differences of the question content in each round of question and answer text in a single stored interaction sample. The intent jump coefficient reflects the frequency of topic migration and the intensity of intent change for the user in the same interaction process.

[0039] Specifically, for a single question in the stored interaction sample, its corresponding semantic difference is determined based on the average of the sub-differences of the question and the sub-differences of the other two questions; the sub-differences of any two questions are determined based on the difference between 1 and the semantic similarity of the two questions. The semantic similarity of the two questions is obtained by mapping the two question texts to sentence vectors through the same semantic coding model and then calculating the cosine similarity. The semantic coding model includes, but is not limited to, SBERT, SimCSE and BERT-whitening. In this embodiment of the invention, for a single storage interaction sample, its corresponding industry sensitivity is determined based on the ratio of the number of associated samples corresponding to that sample to the total number of storage interaction samples. Industry sensitivity reflects the degree of knowledge overlap with users in the same industry. The higher the degree of overlap, the higher the risk of competitive information leakage.

[0040] Among them, the associated sample is the storage interaction sample generated by a user who has a peer relationship with the user corresponding to the storage interaction sample; wherein, the peer relationship is defined as the difference between the user profiles of the two users being less than the preset difference between the user profiles. In this embodiment of the invention, the user profile difference degree is determined based on the difference between 1 and the ratio of the number of identical keywords. The user profile difference degree reflects the degree of similarity between two users in terms of industry characteristics and behavioral patterns.

[0041] The ratio of identical keywords is the ratio of the number of identical keywords in the reference texts of the two users to the preset number of identical keywords. The number of identical keywords in the reference texts of the two users refers to the total number of keywords contained in the intersection of the two keyword sets after the two reference texts are segmented and keywords are extracted. The reference text is the collection of all questions submitted by a single user on the platform within a preset time window, which is the most recent 50 storage periods.

[0042] In this embodiment of the invention, the number of identical keywords reflects the degree of consistency between the business directions of two users. The greater the need to avoid misjudgment and improve the accuracy of peer identification, the larger the value of the preset number of identical keywords. The greater the need to expand the scope of peer association identification and capture potential peer users, the larger the value of the preset user profile difference. Preferably, historical working conditions that meet the usage requirements are obtained and stored. User combinations that have peer associations in the historical working conditions are obtained. The average number of identical keywords for each user combination is 8. The average value of the user profile difference for each user combination is 0.67. These are respectively denoted as the preset number of identical keywords and the preset user profile difference.

[0043] In this embodiment of the invention, the larger the intent jump coefficient, the more frequently the user switches topics in the same interaction process, the wider the scope of sensitive information involved, and the higher the risk of privacy leakage; the larger the industry sensitivity, the higher the degree of knowledge overlap among users in the same industry, and the higher the degree of overlap, the higher the risk of competitive information leakage; therefore, the greater the requirement for storage security accuracy, the smaller the preset intent jump coefficient and preset industry sensitivity values. Preferably, historical working conditions that meet the usage requirements are obtained, and the minimum value of the intent jump coefficient (0.3) and the minimum value of the industry sensitivity (0.1) corresponding to each highly sensitive sample in the historical working conditions are obtained, which are respectively denoted as the preset intent jump coefficient and the preset industry sensitivity.

[0044] Specifically, the clustering and combination analysis module determines the segmentation of highly sensitive samples based on the concentration of sensitive content.

[0045] Specifically, the clustering and combination analysis module determines that no fragmentation is needed for low-sensitivity samples, and records each storage interaction sample that is classified as a low-sensitivity sample as a data segment. In this embodiment of the invention, the process of segmenting based on the concentration of sensitive content includes: When processing a single highly sensitive sample by segmentation, the question and answer texts with a concentration of sensitive content greater than the preset concentration of sensitive content are first recorded as independent data segments. Then, according to the question time from earliest to latest, the question and answer texts of each round that were not recorded as data segments in this highly sensitive sample were segmented and analyzed. When performing segmented analysis on single-round question and answer texts, each round of question and answer texts is recorded into a text combination in turn, and the concentration of sensitive content in the current text combination is calculated after each round is recorded. If the concentration of sensitive content in the current text combination is greater than the preset concentration of sensitive content, then all the question and answer texts that were not recorded in the data paragraphs before this round of question and answer texts will be recorded into a data paragraph. We will continue to segment and analyze the question and answer texts that have not yet been included in the data segments until all question and answer texts are included in the corresponding data segments.

[0046] In this embodiment of the invention, for a number of rounds of question and answer texts, the concentration of sensitive content is determined based on the ratio of the number of different sensitive keywords appearing in the number of rounds of question and answer texts to the total number of different sensitive keywords appearing in the question and answer texts. The sensitive keywords include keywords whose graph connection number is greater than the preset graph connection number, as well as keywords corresponding to newly added nodes; The graph connection number corresponding to a single keyword is the number of nodes directly connected to the node corresponding to that keyword in the initial knowledge graph. The graph connection number reflects the density of association between the entity corresponding to the keyword and surrounding knowledge nodes. Therefore, the larger the value of the graph connection number, the more central the entity corresponding to the keyword is in the knowledge graph, with broader knowledge associations and higher information radiation capabilities. Thus, the greater the need to improve the accuracy of sensitive content identification and avoid misjudging edge nodes as sensitive nodes, the larger the value of the preset graph connection number should be. Preferably, historical working conditions that meet the usage requirements are obtained and stored. The average value of the graph connection number corresponding to each sensitive keyword in the historical working conditions is 9, and this average value is recorded as the preset graph connection number.

[0047] In this embodiment of the invention, when determining cluster combinations based on graph difference degree and user profile difference degree, the graph difference degree of any two data segments in a single cluster combination is greater than a preset graph difference degree, and the user profile difference degree of the corresponding users of the two data segments is greater than a preset user profile difference degree; and the graph difference degree between any data segment outside the cluster combination and at least one data segment in the cluster combination is less than or equal to a preset graph difference degree or the user profile difference degree of the corresponding users of the two data segments is less than or equal to a preset user profile difference degree. Specifically, for any two data segments, the graph difference degree is determined by the difference between 1 and the node number ratio. The graph difference degree reflects the degree of structural difference in the knowledge content carried by the two data segments. The node number ratio is the ratio of the number of identical nodes in the two data segments to the total number of different nodes in the two data segments.

[0048] In this embodiment of the invention, the greater the graph difference, the more significant the difference in knowledge structure between the two data segments, the lower the degree of overlap of the information they carry, and the smaller the risk of sensitive information amplification. Therefore, the greater the need to avoid the risk of sensitive information amplification due to similar knowledge structures, the larger the value of the preset graph difference. Preferably, historical working conditions that meet the usage requirements are obtained and stored. The average value of the graph difference between any two data segments in each cluster combination in the historical working conditions is 0.8, and this average value is recorded as the preset graph difference.

[0049] Please continue reading. Figure 1 and Figure 2 As shown in the embodiment of the present invention, the optional node determination module determines to disable a node if the desensitization time expansion coefficient is greater than the preset desensitization time expansion coefficient or the sensitivity interaction degree is greater than the preset sensitivity interaction degree, and sets the disabling time window duration to the base duration. The optional node determination module determines that nodes whose desensitization time dilation coefficient is less than or equal to the preset desensitization time dilation coefficient and whose sensitivity interaction degree is less than or equal to the preset sensitivity interaction degree do not need to be deactivated.

[0050] In this embodiment of the invention, the desensitization time dilation coefficient is determined based on the ratio of the processing offset difference of a single node to the processing offset benchmark. The desensitization time dilation coefficient reflects the degree of balance of the node's processing capabilities.

[0051] The processing offset difference is determined based on the difference between the maximum and minimum character time consumption index of the cluster combination that has been de-identified in the current storage cycle. The processing offset benchmark is the average character time consumption index of each cluster combination that has been de-identified in the current storage cycle. For a single cluster combination that has undergone de-identification processing on a single node, its character time consumption index is the ratio of time ratio to character length ratio. The character time consumption index reflects the efficiency level of the de-identification processing of the cluster combination on the node. The time ratio is the ratio of the time for de-identification processing of the cluster combination to the preset de-identification time. The character length ratio is the ratio of the number of characters contained in the cluster combination to the preset number of characters. The preset de-identification time is the average time for de-identification processing of all cluster combinations that have undergone de-identification processing on the node in the current storage cycle. The preset number of characters is the average number of characters corresponding to all cluster combinations that have undergone de-identification processing on the node in the current storage cycle. The sensitivity of interaction is determined based on the average of the sub-interactions between a single node and each cluster combination. The sensitivity of interaction reflects the overall interaction tightness between the node and the cluster combination to be processed within the current storage period and the risk of cross-propagation of sensitive information. The sub-interaction of the node with a single cluster combination is the average of the interaction reference values ​​between the cluster combination and each cluster combination in the current storage period in which the node has not yet undergone de-sensitization processing. For any two cluster combinations, their corresponding interaction reference values ​​are positively correlated with both the graph difference deviation and the user profile difference deviation. Specifically, the graph difference deviation is determined by the difference between 1 and the graph difference ratio, and the user profile difference deviation is also determined by the difference between 1 and the user profile difference ratio. The graph difference ratio is the ratio of the graph difference to a preset graph difference, and the user profile difference ratio is the ratio of the user profile difference to a preset user profile difference. In this embodiment, the interaction reference value is obtained by adding the graph difference deviation and the user profile difference deviation together. In this embodiment of the invention, the deactivation time window is the time interval during which a node is prohibited from processing new tasks after the system triggers a deactivation request once it is determined to be deactivated. The base duration of the deactivation time window is set to 5 minutes, and this base duration is combined with the average recovery time of a node during normal faults to serve as the basic time for temporary node isolation. When a node is determined to enter the deactivation time window, the system automatically issues a deactivation request, prohibiting the node from accessing new clustering storage tasks. The deactivation request sets the node as an unavailable node.

[0052] In this embodiment of the invention, the larger the desensitization time dilation coefficient and the higher the sensitivity interaction degree, the worse the balance of node processing capabilities and the higher the sensitivity of node-associated samples. Therefore, the greater the user's demand for the availability of the corresponding node, the smaller the values ​​of the preset desensitization time dilation coefficient and the preset sensitivity interaction degree, so as to reduce the deactivation trigger threshold and protect more nodes in the available state. Preferably, historical working conditions that meet the usage requirements are obtained and stored. The average value of the desensitization time dilation coefficient corresponding to each selectable node in the historical working conditions is 0.6, and the average value of the sensitivity interaction degree is 0.5, which are respectively denoted as the preset desensitization time dilation coefficient and the preset sensitivity interaction degree.

[0053] Please see Figure 3 As shown, it is a flowchart of an embodiment of the present invention for determining whether to reduce the duration of the shutdown time window based on the sensitivity dilution and the intermittent entropy of the shutdown request.

[0054] Specifically, the optional node determination module determines to reduce the duration of the shutdown time window if the sensitive dilution is greater than the preset sensitive dilution and the intermittent entropy of the shutdown request is less than the preset intermittent entropy of the shutdown request.

[0055] Specifically, the optional node determination module determines that it will not adjust the duration of the shutdown time window if the sensitive dilution is less than or equal to the preset sensitive dilution or the intermittent entropy of the shutdown request is greater than or equal to the preset intermittent entropy of the shutdown request.

[0056] In this embodiment of the invention, the sensitivity dilution is determined by the ratio of the total amount of data segments to the number of stored interactive samples. The sensitivity dilution reflects the sparsity of the density of the currently carried data segments. The intermittent entropy of stop request is determined based on the entropy contribution value corresponding to different time interval values. The intermittent entropy of stop request reflects the irregularity and randomness of node stop requests in the time dimension. In this embodiment of the invention, the intermittent entropy of stop request is the negative of the sum of the entropy contribution values ​​corresponding to all different time interval values. Specifically, for a single time interval value, its entropy contribution is the probability of occurrence corresponding to that interval value multiplied by the base-2 logarithm of that probability of occurrence. It should be noted that when the probability of occurrence is zero, its entropy contribution is considered zero. The probability of occurrence corresponding to a single time interval value is the ratio of the number of times that time interval value appears in the time interval sequence to the total number of intervals in the time interval sequence. For a single node, its time interval sequence consists of the time interval between two adjacent stop requests, and the time interval is calculated based on the time when each stop request is triggered by that node.

[0057] It should be noted that when the number of pause requests is less than two, no time interval can be calculated, and the pause request intermittent entropy is directly set to zero.

[0058] In this embodiment of the invention, a higher sensitivity dilution and a lower intermittent entropy of storage requests indicate a sparser distribution of data segments and a more regular behavior of node storage requests, resulting in a more stable overall system operation. Therefore, the greater the user's demand for storage resource utilization efficiency and system operational stability, the smaller the preset sensitivity dilution and the larger the preset intermittent entropy of storage requests. Preferably, historical operating conditions are obtained where storage meets usage requirements and the duration of the downtime window is reduced. The average sensitivity dilution of each storage cycle in these historical operating conditions is 1.6, and the average intermittent entropy of each storage cycle is 0.4, which are respectively denoted as the preset sensitivity dilution and the preset intermittent entropy of storage requests.

[0059] In this embodiment of the invention, the decrease in the duration of the deactivation time window is positively correlated with the sensitive dilution deviation value; wherein, the decrease in the duration of the deactivation time window is determined by multiplying the sensitive dilution deviation value by the baseline duration; the sensitive dilution deviation value is determined by the ratio of the sensitive dilution difference to the preset sensitive dilution; the sensitive dilution difference is the difference between the sensitive dilution and the preset sensitive dilution.

[0060] Please continue reading. Figures 1 to 3 As shown in the embodiment of the present invention, the optional node determination module determines that nodes that are not in the shutdown time window are optional nodes.

[0061] Specifically, the storage execution module allocates optional nodes for each cluster combination based on the sensitivity reference value and the correlation interaction coefficient; The storage execution module determines the allocation priority coefficient of each cluster combination based on the sensitive reference value, and determines the optional nodes of each cluster combination based on the association interaction coefficient. Among them, the allocation priority coefficient of a single cluster combination is positively correlated with the sensitivity reference value, and the optional nodes allocated to a single cluster combination are the optional nodes with the smallest association interaction coefficient with that cluster combination.

[0062] In this embodiment of the invention, the sensitive reference value is the ratio of the reference number of a single cluster combination to the average reference number of each cluster combination. The sensitive reference value reflects the density of sensitive keywords of the cluster combination relative to the average level. The reference number of a single cluster combination is the number of different sensitive keywords appearing in the cluster combination. In this embodiment of the invention, the allocation priority coefficient of a single cluster combination is equal to the sensitivity reference value. It can be understood that the higher the value of the allocation priority coefficient, the earlier the cluster combination should be allocated to an available node for desensitization processing within the current storage cycle.

[0063] In this embodiment of the invention, when determining the optional nodes of each cluster combination based on the association interaction coefficient, the optional node with the smallest association interaction coefficient with the cluster combination is selected for a single cluster combination. Specifically, for a single selectable node and a single cluster combination, the association interaction coefficient is determined based on the average of the interaction reference values ​​between the cluster combination and each reference cluster combination. The reference cluster combinations include all cluster combinations that have not yet been processed by the node, as well as cluster combinations with an allocation priority coefficient greater than the current cluster combination that have been allocated to the node.

[0064] Please see Figure 4 As shown, it is a flowchart of an embodiment of the present invention for determining whether a region is in a sensitive accumulation state based on the risk accumulation coefficient and the search hotspot connectivity.

[0065] Specifically, the storage optimization module determines that it is in a sensitive accumulation state when the risk backlog coefficient is greater than the preset risk backlog coefficient or the search hotspot connectivity is greater than the preset search hotspot connectivity.

[0066] The risk backlog coefficient is determined based on the ratio of the number of loads to the number of references. The risk backlog coefficient is used to reflect the processing and backlog pressure of the current storage cycle relative to the previous adjacent storage cycle. In this embodiment of the invention, the number of loads is determined based on the sum of the number of unprocessed historical cluster combinations of the current storage cycle and the number of cluster combinations of the current storage cycle. The unprocessed historical cluster combinations are the cluster combinations of the previous adjacent storage cycle that have not yet undergone desensitization processing at the end of the current storage cycle. The number of references is determined based on the sum of the number of cluster combinations of the previous adjacent storage cycle and a preset constant. The preset constant is 1 to prevent the denominator from being zero.

[0067] In this embodiment of the invention, a larger risk backlog coefficient indicates a heavier load pressure in the current storage cycle compared to the processing capacity of the previous cycle, and a higher risk of system backlog. Therefore, the higher the requirements for system stability and processing efficiency, the smaller the preset risk backlog coefficient should be. Preferably, historical operating conditions where storage meets usage requirements and is in a sensitive backlog state are obtained, and the average risk backlog coefficient in the historical operating conditions is 1.4. This average value is recorded as the preset risk backlog coefficient.

[0068] The search hotspot connectivity is determined based on the coupling tendency value of each cluster combination. The search hotspot connectivity reflects the degree of association with the current public opinion hotspot. In this embodiment of the invention, the search hotspot connectivity is the average value of the coupling tendency value of each cluster combination. For a single cluster combination, its corresponding coupling tendency value is determined based on the ratio of the number of highly coupled cluster combinations to the number of cluster combinations in the current storage cycle. The coupling tendency value reflects the information coupling strength between the cluster combination and other cluster combinations. In this embodiment of the invention, for a single cluster combination, the number of highly coupled cluster combinations is the number of cluster combinations whose hotspot connection reference value with the cluster combination is less than a preset hotspot connection reference value. The hotspot connection reference value is determined based on the average number of connection edges of each public opinion hotspot keyword corresponding to two cluster combinations. The number of edges connecting a single hot topic keyword is the number of the minimum edges between that hot topic keyword and other hot topic keywords in another cluster combination. The minimum edge between two keywords is the number of edges traversed in the initial knowledge graph when starting from the node corresponding to one keyword and reaching the node corresponding to the other keyword along the shortest path. If a single hot topic keyword exists in both cluster combinations, the number of minimum edges is 0. Hot topic keywords are those that appear in the question content of interactive samples, and the number of interactive samples in which the keyword appears is greater than the preset number of interactive samples. Preferably, historical working conditions that meet the usage requirements and are in a sensitive accumulation state are obtained and stored. The number of interactive samples in which each hot topic keyword appears in the corresponding storage period is counted. The average number of interactive samples corresponding to each hot topic keyword is 67, and this average value is determined as the preset number of interactive samples.

[0069] In this embodiment of the invention, the search hotspot connectivity is used to reflect the average level of coupling between cluster combinations. It is understood that the higher the coupling tendency value of each cluster combination, the more public opinion hotspot information is shared between the cluster combinations, and the shorter the public opinion propagation path. This easily leads to a closer correlation between public opinion hotspot information between cluster combinations, resulting in a higher risk of public opinion hotspot accumulation when the system processes interactive samples, leading to excessive concentration and cross-propagation of public opinion hotspot information, thus increasing the risk of user privacy leakage. Therefore, the greater the demand for public opinion information isolation capabilities and user privacy protection, the smaller the preset search hotspot connectivity value. Preferably, historical working conditions that meet usage requirements and are in a sensitive accumulation state are obtained and stored. The average value of the search hotspot connectivity in the historical working conditions is 0.42, and this average value is recorded as the preset search hotspot connectivity.

[0070] The risk backlog coefficient reflects the processing and backlog pressure of the current storage cycle relative to the previous adjacent storage cycle. The search hotspot connectivity reflects the degree of correlation between each cluster combination and the current public opinion hotspot. Therefore, the storage optimization module determines whether it is in a sensitive accumulation state by comparing the risk backlog coefficient and the search hotspot connectivity with the corresponding preset thresholds. It can comprehensively evaluate the system status from two dimensions: storage processing load and public opinion information coupling, avoiding the misjudgment or omission caused by a single indicator.

[0071] Please continue reading. Figures 1 to 4 As shown in this embodiment of the invention, the storage optimization module performs sensitive combination delayed storage; The storage optimization module performs delayed storage for each sensitive combination, and the delay time corresponding to each sensitive combination is positively correlated with the risk assessment value.

[0072] Sensitive clusters are clusters whose sensitive reference values ​​are greater than preset sensitive reference values. A sensitive reference value is used to reflect the sensitivity level of cluster combinations. It is understood that when performing delayed storage of sensitive combinations, cluster combinations with higher sensitive reference values ​​contain more sensitive keywords, meaning the combination carries more densely packed privacy information. Therefore, the greater the demand for system desensitization efficiency and sensitive information control capabilities, the smaller the preset sensitive reference value should be to avoid the security problem of cross-leakage of sensitive information caused by simultaneous transmission to optional nodes. Preferably, historical operating conditions where storage meets usage requirements and delayed storage of sensitive combinations is performed are obtained, and the average value of the sensitive reference values ​​corresponding to each sensitive combination in the historical operating conditions is calculated to be 0.7. This average value is recorded as the preset sensitive reference value.

[0073] The delay time is determined based on the product of the baseline delay time and the contribution coefficient. In this embodiment, if the risk assessment value is greater than the preset risk assessment value, the contribution coefficient is set as the ratio of the risk assessment value to the preset risk assessment value; if the risk assessment value is less than or equal to the preset risk assessment value, the contribution coefficient is set to 1. The baseline delay time is set to 10 minutes. The greater the need for risk control of sensitive system data and prevention of centralized leakage, the larger the value of the baseline delay time. In this embodiment, the risk assessment value is the sum of the first risk index and the second risk index. The first risk index is the value after scaling the risk backlog coefficient ratio with the first contribution parameter, and the second risk index is the value after scaling the search hotspot connectivity ratio with the second contribution parameter. The risk backlog coefficient ratio is the ratio of the risk backlog coefficient to the preset risk backlog coefficient, the search hotspot connectivity ratio is the ratio of the search hotspot connectivity to the preset search hotspot connectivity, and the sum of the contribution parameters is 1. Preferably, the risk backlog coefficient ratio and the search hotspot connectivity ratio both reflect the risk level of the current node storage in the system from different perspectives. For the contribution parameter, the greater the user's sensitivity to the numerical item, the larger the contribution parameter. The first contribution parameter is 0.5, and the second contribution parameter is 0.5.

[0074] In this embodiment of the invention, the risk assessment value is used to reflect the degree of comprehensive accumulated risk and privacy leakage threat currently faced by the system. It is understood that a higher risk assessment value indicates a more prominent risk in both storage processing load and public opinion information coupling, necessitating the triggering of sensitive accumulation state determination and corresponding delayed storage processing. Therefore, the greater the demand for system risk prevention and control capabilities and privacy protection security, the smaller the preset risk assessment value should be. Preferably, historical operating conditions where storage meets usage requirements and sensitive combination delayed storage is obtained, and the average risk assessment value of these historical operating conditions is calculated to be 0.65. This average value is recorded as the preset risk assessment value.

[0075] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A storage system for data encryption and privacy protection in an AI-powered business social platform, characterized in that, include: The storage sample selection module is used to determine the storage interaction samples based on the knowledge gain coefficient and the question depth value; The clustering and combination analysis module is used to determine the category of stored interaction samples based on intent jump coefficient and industry sensitivity, and to determine whether to perform segmentation based on the concentration of sensitive content based on the category of stored interaction samples, and to determine clustering combination based on graph difference and user profile difference. The optional node determination module is used to determine whether to disable a node based on the desensitization time inflation coefficient and the sensitivity of the interaction, and to determine whether to reduce the duration of the disabling time window based on the sensitivity dilution and the intermittent entropy of the disabling request, and to determine the optional node based on whether it is in the disabling time window. The storage execution module is used to allocate optional nodes for each cluster combination based on the sensitivity reference value and the correlation interaction coefficient; The storage optimization module is used to determine whether the system is in a sensitive accumulation state based on the risk backlog coefficient and the search hotspot connectivity, and to perform sensitive combination delayed storage when the system is in a sensitive accumulation state.

2. The storage system for data encryption and privacy protection for AI business social platforms according to claim 1, characterized in that, The storage sample selection module determines that interactive samples with a knowledge gain coefficient greater than a preset knowledge gain coefficient and a question depth value greater than a preset question depth value are stored interactive samples.

3. The storage system for data encryption and privacy protection for AI business social platforms according to claim 2, characterized in that, The clustering and combination analysis module determines that stored interaction samples with an intent jump coefficient greater than or equal to a preset intent jump coefficient or a peer sensitivity greater than or equal to a preset peer sensitivity are highly sensitive samples. The clustering and combination analysis module determines that stored interaction samples with an intent jump coefficient less than a preset intent jump coefficient and a peer sensitivity less than a preset peer sensitivity are low-sensitivity samples.

4. The storage system for data encryption and privacy protection for AI business social platforms according to claim 3, characterized in that, The clustering and combination analysis module determines the segmentation of highly sensitive samples based on the concentration of sensitive content.

5. The storage system for data encryption and privacy protection for AI business social platforms according to claim 1, characterized in that, The optional node determination module determines to disable nodes whose desensitization time expansion coefficient is greater than the preset desensitization time expansion coefficient or whose sensitivity interaction degree is greater than the preset sensitivity interaction degree, and sets the disabling time window duration to the base duration. The optional node determination module determines that nodes whose desensitization time dilation coefficient is less than or equal to the preset desensitization time dilation coefficient and whose sensitivity interaction degree is less than or equal to the preset sensitivity interaction degree do not need to be deactivated.

6. The storage system for data encryption and privacy protection for AI business social platforms according to claim 5, characterized in that, The optional node determination module determines to reduce the duration of the shutdown time window if the sensitive dilution is greater than the preset sensitive dilution and the intermittent entropy of the shutdown request is less than the preset intermittent entropy of the shutdown request.

7. The storage system for data encryption and privacy protection for AI business social platforms according to claim 6, characterized in that, The optional node determination module determines that nodes that are not in the shutdown time window are optional nodes.

8. The storage system for data encryption and privacy protection for an AI business social platform according to claim 1, characterized in that, The storage execution module allocates selectable nodes for each cluster combination based on the sensitivity reference value and the correlation interaction coefficient. The storage execution module determines the allocation priority coefficient of each cluster combination based on the sensitive reference value, and determines the optional nodes of each cluster combination based on the association interaction coefficient. Among them, the allocation priority coefficient of a single cluster combination is positively correlated with the sensitivity reference value, and the optional nodes allocated to a single cluster combination are the optional nodes with the smallest association interaction coefficient with that cluster combination.

9. The storage system for data encryption and privacy protection for an AI business social platform according to claim 8, characterized in that, The storage optimization module determines that it is in a sensitive accumulation state when the risk backlog coefficient is greater than the preset risk backlog coefficient or the search hotspot connectivity is greater than the preset search hotspot connectivity.

10. The storage system for data encryption and privacy protection for an AI business social platform according to claim 9, characterized in that, The storage optimization module performs sensitive combined delayed storage; The storage optimization module performs delayed storage for each sensitive combination, and the delay time corresponding to each sensitive combination is positively correlated with the risk assessment value.

Images