Blockchain-based edge llm trusted knowledge filtering, sharing, and caching method
By introducing multi-dimensional quality assessment and a delay-trust trade-off strategy in Large Language Model (LLM) knowledge sharing using blockchain technology, the problems of uneven data quality and real-time reliability are solved, and efficient and trustworthy knowledge sharing and cache management are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUBEI UNIV OF TECH
- Filing Date
- 2026-03-27
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies in the process of knowledge sharing of large language models (LLM) suffer from inconsistent data quality, lack of multi-dimensional verification mechanisms, imperfect reputation management of participating entities, and insufficient consideration of the balance between latency and credibility in knowledge screening, making it difficult to meet the dual requirements of real-time performance and reliability.
By leveraging blockchain technology to achieve data traceability and trust transfer, and employing a multi-dimensional quality assessment model and a delay-trust trade-off strategy, combined with reputation mechanisms and multi-dimensional quality assessment models, data filtering and caching management are carried out to ensure data quality and real-time performance.
It enables high-quality, low-latency knowledge sharing, ensures the credibility and traceability of data, and improves the real-time inference efficiency and data sharing security of LLM.
Smart Images
Figure CN122332573A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence and blockchain technology, and more specifically, to a blockchain-based edge LLM trusted knowledge filtering, sharing and caching method. Background Technology
[0002] With the rapid iteration of artificial intelligence technology, Large Language Models (LLMs), with their flexible parameter adjustment and scenario adaptation capabilities, have been widely applied in various fields such as intelligent interaction, data analysis, and knowledge services. Knowledge sharing, as a core component for improving the inference accuracy and dynamic update efficiency of LLMs, has become a key research direction in the industry. To address the problems of insufficient data quality, excessively long response times, and insufficient credibility in traditional knowledge sharing processes, blockchain technology, with its distributed notarization and immutability characteristics, holds promise for introduction into large model knowledge sharing scenarios.
[0003] However, existing technologies still have many pain points that urgently need to be addressed: First, the quality of data varies greatly during knowledge sharing, and there is a lack of multi-dimensional verification mechanisms. Relying solely on a single score or simple logical rules to filter knowledge makes it difficult to guarantee the factual consistency, task relevance, and security of shared data, thus limiting the effectiveness of LLM updates. Second, the reputation management system of participating entities is imperfect, and a two-way reputation update mechanism between data providers and verifiers has not been established, which cannot effectively constrain behaviors such as malicious provision of false data and distorted verification results, affecting the credibility of the knowledge sharing ecosystem. Third, knowledge screening does not fully consider the balance between latency and credibility. Existing solutions mostly focus on single-dimensional optimization, which is difficult to meet the dual requirements of real-time performance and reliability for LLM inference.
[0004] While some existing technologies attempt to achieve knowledge graph integration, large-model collaborative verification, or caching optimization through blockchain, they struggle to simultaneously address the comprehensive technical challenges of data credibility, screening accuracy, inference efficiency, and process traceability. Therefore, there is an urgent need for an LLM knowledge-sharing method within a blockchain scenario that integrates multi-dimensional verification, bidirectional reputation updates, latency-reliability balanced screening, and intelligent hierarchical caching. This method aims to overcome existing technological bottlenecks and achieve high-quality, low-latency knowledge sharing. Summary of the Invention
[0005] This invention addresses the technical problems existing in LLM edge inference data sharing, such as data silos, knowledge conflicts, imbalance between real-time performance and reliability, and low data quality. It provides a blockchain-based method for edge LLM trusted knowledge filtering, sharing, and caching. The method uses blockchain to achieve data traceability and trust transfer, optimizes the reputation mechanism and multi-dimensional quality assessment model to ensure data quality, and balances the real-time performance and reliability requirements through a delay-reliability trade-off strategy. Ultimately, it supports continuous and reliable data acquisition and model updates for LLM.
[0006] According to a first aspect of the present invention, a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method is provided, comprising: S1, the data provider node listens for knowledge requests on the blockchain and uploads the metadata of the data to be shared and the storage certificate to the blockchain for evidence storage. S2, the validator node obtains the data to be shared from the blockchain and evaluates it from multiple preset dimensions to generate a verification score; S3, based on the generated verification score, the blockchain network updates the reputation of the data items provided by the data provider node and the reputation of the verifier node itself through a preset reputation calculation model; S4. Based on the preset constraints, combined with the updated data item reputation, the multi-dimensional quality assessment results of the data items, and the trade-off function composed of data latency and credibility, a set of data items that meet the preset constraints is selected from all the data to be shared. S5, the selected set of data items is transmitted from the corresponding data provider node to the cache node for use by the large language model, and the result summary of this screening is stored on the blockchain for evidence. S6, the cache node performs hierarchical storage and dynamic management of the received set of data items.
[0007] Based on the above technical solution, the present invention can also be improved as follows.
[0008] Optionally, in step S3, the reputation of the data item Calculated using the following formula:
[0009] in, To update the weights (0 < <1), This is a credibility penalty coefficient. To verify the weighted mean of the scores, For data credibility, The maximum confidence threshold; The weighted average of the verification scores The calculation model is as follows:
[0010] in, For data items p The weighted average of the verification scores at time t. For the i-th validator, the data item p of t Verify scores at all times. For the i-th verifier tReputation at all times For validator nodes, V This is the set of validator nodes.
[0011] Optionally, the data credibility For all validator nodes, pair data items p The standard deviation of the verification scores is calculated as follows:
[0012] in, The number of validators.
[0013] Optionally, in step S3, the reputation of the validator node Calculated using the following model:
[0014] in, For the i-th verifier Reputation at all times For validator nodes, For validator nodes Scoring consistency bias, It is used to measure the consistency of its score with the scores of other validator nodes.
[0015] Optionally, in step S4, the multi-dimensional quality assessment results of the data items It is a weighted sum of static quality indicators, statistical quality indicators, and task relevance indicators; The static quality index is calculated based on the perplexity of the data item; the statistical quality index is a weighted sum of semantic conflict and redundancy; and the task relevance index is obtained by calculating the matching score between the data item and the current query.
[0016] Optionally, in step S4, the trade-off function consisting of data latency and reliability... , represented as:
[0017] in, and These are the delay weight and the credibility weight, respectively. The preset maximum allowable delay, For data items p Total delay, This includes verification delays, blockchain evidence storage delays, and data transmission delays.
[0018] Optionally, in step S4, the preset constraints include at least one of the following: binary constraints for accepting or not accepting data items, delay constraints, data credibility constraints, and constraints on the maximum number of data items to accept.
[0019] Optionally, in step S4, the set of data items that meet preset constraints is selected from all the data to be shared. The optimization objective of the data selection is to maximize the total utility of the selected set of data items, wherein the utility of a single data item is... This is the sum of the reputation, multi-dimensional quality assessment results, and the compromise function value for the data item.
[0020] Optionally, step S6 includes: The time decay access frequency, data item utility and storage overhead corresponding to each data item in the data item set are weighted and combined to obtain the cache evaluation value corresponding to each data item. Based on the cache evaluation value, all data items in the data item set are sorted, and according to the preset Top-K rule, the top K1 data items with the highest cache evaluation value are stored in the hot cache, the next K2 data items are stored in the warm cache, and the remaining data items are stored in the cold cache. When new data items need to be stored in the cache or the cache space is insufficient, the cache evaluation value of the relevant data items is recalculated, and the data item sorting and storage level are dynamically adjusted based on the updated cache evaluation value.
[0021] According to a second aspect of the present invention, a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching system is provided, comprising: The data provider module is used to listen for knowledge requests on the blockchain network and to store the metadata and storage certificates of the data to be shared on the blockchain. The validator module is used to obtain the data to be shared from the blockchain network, evaluate the data to be shared from multiple preset dimensions, and generate a verification score; The reputation management module, deployed in the blockchain network, is used to update the reputation of data items and the reputation of validator nodes based on the verification score and through a preset reputation calculation model. The data filtering module is used to filter out a set of data items that meet the conditions from all the data to be shared, based on preset constraints, the updated reputation of data items, the multi-dimensional quality assessment results of data items, and a trade-off function composed of data latency and credibility. The data sharing and evidence storage module is used to transfer the selected set of data items from the data provider module to the cache module, and to store the summary of the selection results on the blockchain for evidence storage. The cache management module is used to perform hierarchical storage and dynamic management of the received data item set for use by the edge large language model.
[0022] According to a third aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the processor is configured to implement the steps of the above-described blockchain-based edge LLM trusted knowledge filtering, sharing and caching method when executing a computer management program stored in the memory.
[0023] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer management program is stored, wherein when executed by a processor, the computer management program implements the steps of the above-described blockchain-based edge LLM trusted knowledge filtering, sharing and caching method.
[0024] This invention provides a blockchain-based method, system, electronic device, and storage medium for edge LLM trusted knowledge filtering, sharing, and caching. By integrating data on-chain notarization, multi-dimensional data verification, bidirectional reputation updates combining verification scores and trustworthiness penalties, reputation fusion, multi-dimensional quality assessment, and a constraint-based filtering using a delay-trustworthiness trade-off function, and finally, hierarchical caching management, into a coherent process, it fundamentally improves the overall efficiency of edge LLM knowledge sharing. The specific beneficial technical effects of this invention are reflected in: 1. Solving the problem of data silos: By building a decentralized data sharing framework through blockchain, information barriers between edge nodes are broken down, enabling unified access and trusted sharing of multi-source data. At the same time, a data traceability mechanism ensures that the source of data is transparent and verifiable.
[0025] 2. Mitigating dynamic knowledge conflict: By calculating semantic conflict and redundancy in the multi-dimensional data quality assessment model, low-conflict and low-redundancy data can be effectively identified and filtered, reducing the risk of catastrophic forgetting and improving the robustness of the model.
[0026] 3. Balancing real-time performance and reliability: Design a latency-reliability trade-off function that comprehensively considers the total latency of data transmission, verification, and blockchain notarization. While meeting the real-time inference requirements of DLLM, it ensures data reliability through reliability constraints.
[0027] 4. Ensure data quality: Construct a multi-dimensional evaluation model that includes static quality, statistical quality, and task relevance, and combine it with a reputation mechanism to incentivize high-quality data contributions, effectively addressing the spatiotemporal heterogeneity of edge data.
[0028] 5. Achieve trustworthiness and auditability: The immutability of blockchain ensures that the entire process of data on-chaining, verification, filtering, and updating is traceable and auditable, solving cross-entity trust issues and improving the security and reliability of data sharing. Attached Figure Description
[0029] Figure 1A flowchart of a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method is provided for this invention. Figure 2 A flowchart of a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method for one embodiment is provided. Figure 3 A schematic diagram of a multi-level cache provided for one embodiment; Figure 4 A block diagram of a blockchain-based edge LLM trusted knowledge filtering, sharing and caching system is provided for this invention; Figure 5 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention; Figure 6 This is a schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. Detailed Implementation
[0030] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0031] Figure 1 This invention provides a flowchart of a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method. Figure 2 A flowchart illustrating a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method for a specific embodiment is provided, combined with... Figure 1 and Figure 2 As shown, the method includes steps S1 to S6.
[0032] S1, the data provider node listens for knowledge requests on the blockchain and uploads the metadata and storage certificates of the data to be shared to the blockchain for notarization.
[0033] In this step, the data sharing process is initiated. Specifically, the data provider node deployed at the edge continuously listens for knowledge requests initiated by other dynamic LLMs on the blockchain network, such as inference data requests. Upon detecting a relevant request, the edge node, as the data provider, prepares all the knowledge data it possesses locally that may be related to the request. To ensure the traceability of the data source and the transparency of the sharing process, the node does not directly upload the raw data itself. Instead, it packages and submits key information such as the identification information of the data to be shared, including its data identity, metadata description, generation context information, specific data storage location on the edge node, and the node's own reputation credentials, to the blockchain network for notarization. This completes the registration and disclosure of the knowledge data, laying the foundation for subsequent verification and screening.
[0034] S2, the validator node obtains the data to be shared from the blockchain and evaluates it from multiple preset dimensions to generate a verification score.
[0035] This step assesses the quality and credibility of the data disclosed on the blockchain. Validator nodes in the system retrieve data storage information submitted by the data provider from the blockchain, and then use this information to locate and obtain the complete data to be verified. Subsequently, each validator node independently runs a pre-defined evaluation model, comprehensively scoring the data from multiple core dimensions. These dimensions typically include the factual consistency of the data's content, its relevance to the current LLM inference task, and the security of the data itself. Each validator generates a comprehensive verification score covering these multiple dimensions. This score represents the validator's quantitative judgment of the data's quality and is recorded on the blockchain for later use.
[0036] S3, based on the generated verification score, the blockchain network updates the reputation of the data items provided by the data provider node and the reputation of the verifier node itself through a preset reputation calculation model.
[0037] The purpose of this step is to establish and maintain a dynamic, two-way reputation system to incentivize the provision of high-quality data and impartial verification. After collecting all validators' ratings for a given data item, the blockchain network initiates a reputation update calculation. For the data item itself, its reputation value is updated based on the weighted average of all validator ratings, but a penalty mechanism is introduced: if validators' ratings differ significantly, indicating low data credibility, their reputation will be reduced. For validator nodes, their reputation depends on the consistency of their ratings with those of other validators; if their ratings consistently deviate significantly from those of other high-reputation validators, their reputation will decline. By linking data item reputation with validator reputation and dynamically adjusting it, this approach encourages participants to contribute authentic, high-quality data and conduct prudent verification from a game-theoretic perspective.
[0038] S4. Based on preset constraints, such as delay constraints, credibility constraints, and adoption quantity constraints, combined with the updated data item reputation, the multi-dimensional quality assessment results of the data items, and the trade-off function composed of data delay and credibility, a set of data items that meet the preset constraints is selected from all the data to be shared.
[0039] This step primarily involves selecting the optimal knowledge set under multiple constraints. The system first calculates a comprehensive utility value for each candidate data item. This utility value consists of three parts: the updated reputation of the data item, a multi-dimensional quality assessment score calculated based on static quality, statistical quality, and task relevance, and a crucial latency-credibility trade-off function. This trade-off function quantifies the trade-off between the total latency required to acquire the data and its credibility. Based on this, the system aims to maximize the total comprehensive utility of all adopted data items while satisfying a series of pre-defined hard constraints, including that the total latency of each data item cannot exceed a threshold, its credibility must be higher than a minimum requirement, and there is an upper limit to the total number of data items ultimately adopted. By solving this constrained optimization problem, the system can accurately select a subset of data items that achieves the best balance between real-time performance, reliability, and quality.
[0040] S5, the set of data items that meet the preset constraints is transmitted from the corresponding data provider node to the cache node so that the large language model can call it from the cache, and the result summary of this screening (e.g., selected knowledge list, summary of each score, validator signature) is stored on the blockchain for evidence.
[0041] This step delivers and notarizes the screening results. After the optimal set of data items selected in step S4 is determined, the system notifies the corresponding data provider (edge node). The notified edge node then transmits the selected data items from its local storage to the designated cache node for direct reading and use by the LLM, thereby completing the substantial sharing of knowledge. To ensure the immutability and auditability of the entire screening process, a summary of the key results of this screening, including the final list of selected data items, the scores and utility value summaries of each item, and the digital signature of the verifier, is simultaneously recorded on the blockchain as a new transaction. This creates a complete and trustworthy closed-loop record from data disclosure, verification, screening to final adoption.
[0042] S6, combined Figure 3 As shown, the cache node performs hierarchical storage and dynamic management on the received set of data items, and calculates cache evaluation values and dynamic update strategies to update the cache.
[0043] This step is responsible for the efficient organization and management of shared knowledge to optimize LLM access performance. Upon receiving the filtered set of data items, the cache node does not simply store them; instead, it calculates a dynamic cache evaluation value for each selected data item. For example, this evaluation value considers the data item's recent access frequency, its quality and reputational utility, and storage overhead. Subsequently, the caching system sorts all data items in the set based on their evaluation values and implements tiered storage according to a preset strategy. For instance, the "hot" knowledge with the highest evaluation value is stored in memory as a hot cache, the next highest is stored on disk, and others are stored in more economical cold storage. When new data is added or cache space is insufficient, the evaluation values are recalculated and the data's tiered position is adjusted, thereby achieving intelligent and dynamic updates to the cached content and ensuring that LLM can efficiently access the most relevant and highest-quality knowledge.
[0044] Understandably, given the shortcomings in the background technology, this invention proposes a blockchain-based method for edge LLM trusted knowledge filtering, sharing, and caching. This method integrates data on-chain notarization, multi-dimensional data verification, bidirectional reputation updates combining verification scores and trustworthiness penalties, reputation fusion, multi-dimensional quality assessment, and a constrained filtering method using a delay-trustworthiness trade-off function, along with final hierarchical caching management, into a coherent process, fundamentally improving the overall efficiency of edge LLM knowledge sharing.
[0045] Specifically, firstly, data is verified and stored on the blockchain based on the dimensions of factual consistency, task relevance, and security, ensuring the credibility and traceability of the shared knowledge base. Secondly, by simultaneously calculating the reputation of data items and the reputation of verifiers, and introducing data credibility as a penalty factor, a two-way constraint is constructed to incentivize high-quality data contributions and fair verification, effectively curbing malicious behavior. Then, in the data screening stage, not only are the static quality, statistical quality, and task relevance of the data comprehensively considered, but the total latency caused by data transmission, verification, and blockchain storage is innovatively incorporated into a trade-off function for optimization. This allows for the selection of the data set with the highest overall utility while meeting preset latency, credibility, and quantity constraints, thus directly balancing the dual requirements of real-time performance and reliability for LLM inference at the decision-making level. Finally, by implementing intelligent hierarchical caching of adopted data based on access frequency, utility, and storage overhead, the efficiency of knowledge acquisition and retrieval is further optimized. Overall, through the aforementioned interconnected technical steps, this solution systematically addresses the issues of uncontrollable data quality, lack of trust among participants, and singular optimization objectives under resource constraints, thereby achieving synergistic enhancements in the credibility, quality, efficiency, and controllability of knowledge sharing.
[0046] Based on the above technical solutions, the embodiments of the present invention can be further improved as follows.
[0047] In one possible implementation, in step S3, the reputation of the data item Calculated using the following formula:
[0048] in, To update the weights (0 < <1), This is a credibility penalty coefficient. To verify the weighted mean of the scores, For data credibility, This is the maximum confidence threshold.
[0049] In this embodiment, the weighted average of the verification scores The calculation model is as follows:
[0050] in, For data items p The weighted average of the verification scores at time t. For the i-th validator, the data item p of t Verify scores at all times. For the i-th verifier t Reputation at all times For validator nodes, V This is the set of validator nodes.
[0051] Furthermore, the reputation of the validator node Calculated using the following model:
[0052] in, For the i-th verifier Reputation at all times For validator nodes, For validator nodes Scoring consistency bias, It is used to measure the consistency of its score with the scores of other validator nodes.
[0053] Among them, validator rating consistency bias The calculation model is as follows:
[0054] In this embodiment, For the number of validators, To achieve the maximum verification score, To achieve the minimum verification score, This represents the maximum reputation value for the validator.
[0055] In this embodiment, the data credibility For all validator nodes, pair data items p The standard deviation of the verification scores is calculated as follows:
[0056] That is, the standard deviation of the validation scores. The number of validators.
[0057] In this embodiment, a two-way dynamic reputation mechanism is constructed by associating and cyclically constraining data item reputation with verifier reputation, jointly improving the credibility and data quality of knowledge sharing. On one hand, the data item reputation model introduces an average score weighted by verifier reputation as a positive incentive and incorporates a penalty factor based on rating dispersion (i.e., data credibility). This effectively identifies and suppresses low-quality or controversial data, incentivizing data providers to continuously contribute high-quality, highly consistent knowledge. On the other hand, the verifier reputation model dynamically adjusts a verifier's reputation by calculating the consistency deviation between their rating and the mainstream community rating. This reduces the reputation of verifiers who consistently provide abnormal ratings that deviate from consensus, thus weakening the weight of their subsequent ratings. This constrains malicious or inaccurate behavior by verifiers and encourages them to conduct objective and fair evaluations. These two models are coupled and mutually restraining: high-quality data relies on fair ratings from high-reputation verifiers to accumulate reputation, while a verifier's high reputation needs to be maintained by providing consensus ratings for high-quality data. This embodiment filters out high-quality data and reliable verification nodes for the entire sharing process, laying a credible foundation for the sustainable operation of the system.
[0058] In one possible implementation, in step S4, the multi-dimensional quality assessment results of the data items are... It is a weighted sum of static quality indicators, statistical quality indicators, and task relevance indicators; The static quality index is calculated based on the perplexity of the data item; the statistical quality index is a weighted sum of semantic conflict and redundancy; and the task relevance index is obtained by calculating the matching score between the data item and the current query.
[0059] For example, multidimensional data quality The calculation model is as follows:
[0060] in, , , These are static quality weights, statistical quality weights, and task-related weights, respectively. These are static quality indicators. , For data items p The level of confusion, The maximum perplexity threshold; For statistical quality indicators, , For balance coefficient, For semantic conflict degree, For redundancy; , For data items p eigenvectors, As the average vector of the reference set, (Here, i∈H, H is the reference data set); (Here i=1 to m, For data items p Similarity with the top-m existing knowledge (calculated using the FAISS tool). As a task relevance indicator, Calculate data items using the Cross-encoder model p and t Time-based reasoning query Match score, To query the target element.
[0061] Understandably, the multi-dimensional quality assessment model established in this embodiment integrates three types of indicators—static quality, statistical quality, and task relevance—to achieve a comprehensive quantitative evaluation of the intrinsic quality of data, knowledge system compatibility, and scenario adaptability, thereby improving the accuracy and practicality of knowledge selection. Specifically, firstly, it uses a static quality indicator based on perplexity to filter low-quality, chaotic raw data from the perspectives of text fluency and grammatical correctness. Secondly, it uses the semantic conflict and redundancy calculations in the statistical quality indicators to quantitatively assess the consistency between new data and the existing knowledge base, effectively avoiding the risk of introducing contradictory information or duplicate content that could lead to model cognitive confusion or decreased efficiency. Finally, the introduced task relevance indicator, through deep calculation of the matching score between the data and the current specific query intent, ensures that the selected knowledge is highly compatible with the real-time reasoning tasks faced by the edge large language model. This comprehensive evaluation system overcomes the limitations of single-dimensional scoring (such as only considering accuracy or only considering relevance), and can jointly guarantee the high quality, consistency, and usability of the final adopted data from multiple technical levels, providing comprehensive and reliable quality input for subsequent optimization and selection decisions.
[0062] In one possible implementation, in step S4, the trade-off function consisting of data latency and reliability... , represented as:
[0063] in, and These are the delay weight and the credibility weight, respectively. The preset maximum allowable delay, For data items p Total delay, This includes verification delays, blockchain evidence storage delays, and data transmission delays.
[0064] The total delay is expressed as:
[0065] in, To verify the delay, , To verify the start time, To verify the end time; Delay in blockchain evidence storage , The block generation cycle; For data transmission delay, RTT stands for Round Trip Time.
[0066] In this embodiment, a latency-reliability trade-off function is used to unify and optimize the often conflicting objectives of total data acquisition latency and data reliability. This achieves intelligent dynamic decision-making balance in edge LLM knowledge sharing scenarios where both real-time performance and reliability are critical. Specifically, this latency-reliability trade-off function explicitly incorporates the total latency of data items (covering the time spent on verification, on-chain notarization, and transmission) and their reliability (i.e., the consistency of scores from various verifiers) into the calculation framework, and allows for adjustments to latency weights. With credibility weight To flexibly adapt to the preferences of different application scenarios.
[0067] This embodiment enables the calculation of a balanced value between speed and reliability for each candidate data item during the final screening process, thus avoiding the shortcomings of traditional solutions that often compromise on one aspect while neglecting another. Ultimately, this embodiment ensures that, under preset latency constraints, more reliable and less controversial knowledge data is prioritized for screening, or, provided that minimum reliability requirements are met, data with lower latency is prioritized for adoption. This achieves optimal control over the efficiency and quality of knowledge acquisition in resource-constrained edge environments.
[0068] In one possible implementation, in step S4, the preset constraints include at least one of the following: binary constraints for accepting or not accepting data items, delay constraints, data credibility constraints, and constraints on the maximum number of data items to accept.
[0069] Specifically, the preset constraints include: Binary constraints: ， =1 Indicates the adoption of data items p， =0 This indicates that the suggestion is not adopted; Delay constraints: ( (This is the preset maximum allowable delay). Credibility constraints: ( ρ (Preset minimum confidence threshold); Adoption quantity constraints: （ (Maximum number of data items accepted).
[0070] In this embodiment, the optimization objective of data filtering is to maximize the total utility of the selected set of data items, wherein the utility of a single data item is... This is the sum of the reputation, multi-dimensional quality assessment results, and the compromise function value for the data item.
[0071] Specifically, the optimization objective for data filtering is expressed as:
[0072] in, That is, the optimal set of data items is determined by maximizing the sum of the reputation, quality, and delay-credibility trade-off of the adopted data items.
[0073] Understandably, in this embodiment, step S4 transforms the complex data screening problem into a decision-making process seeking the optimal solution within strict boundaries by combining explicit engineering constraints with calculable optimization objectives. This ensures the practicality, efficiency, and interpretability of the screening results. The set of multiple constraints (such as upper limits for delay, lower limits for credibility, and maximum number of adoptions) defines a clear feasible solution space for the entire screening system, forcibly guaranteeing that the final adopted data set must simultaneously meet the requirements of real-time performance, reliability, and resource feasibility. The utility maximization optimization objective plays a core decision-making role. The optimization objective function adds the reputation of a data item, the results of multi-dimensional quality assessment, and the delay-credibility trade-off as the utility of a single data item, and solves for maximizing the total utility of all adopted data items. This embodiment enables the intelligent comparison and weighing of the comprehensive value of different data items within the constraints, automatically selecting the data combination with the highest comprehensive score in quality, credibility, real-time performance, and reputation. This allows for the accurate and efficient delivery of the highest quality knowledge supplement to LLM in resource-constrained edge environments.
[0074] In one possible implementation, refer to Figure 3 As shown, step S6 includes: The time decay access frequency, data item utility and storage overhead corresponding to each data item in the data item set are weighted and combined to obtain the cache evaluation value corresponding to each data item. Based on the cache evaluation value, all data items in the data item set are sorted, and according to the preset Top-K rule, the top K1 data items with the highest cache evaluation value are stored in the hot cache, the next K2 data items are stored in the warm cache, and the remaining data items are stored in the cold cache. When new data items need to be stored in the cache or the cache space is insufficient, the cache evaluation value of the relevant data items is recalculated, and the data item sorting and storage level are dynamically adjusted based on the updated cache evaluation value.
[0075] For example, the calculation model for cache evaluation values is as follows:
[0076] in, , , These are the weighting coefficients.
[0077] The time-decayed access frequency reflects the recent usage popularity of knowledge items, and the formula is: , λ This is a decay coefficient, ensuring that older access records have lower weights.
[0078] The reputation of a knowledge item is derived from the ratings of blockchain validators, representing its quality and authority.
[0079] The storage overhead for knowledge items reflects the physical resource costs they occupy.
[0080] In this embodiment, the cache adopts the Top-K filtering rule, and writes data into the cold cache (IPFS storage method), warm cache (stored on disk), or hot cache (stored in memory) respectively. The hot cache data keeps the latest knowledge state and can be quickly returned to the LLM after the cache hit of the LLM inference request.
[0081] Understandably, this embodiment improves the efficiency of knowledge access and storage resource utilization in edge environments through a tiered caching and dynamic management scheme. The scheme first calculates a cache evaluation value for each data item, comprehensively considering its time-decay access frequency, its intrinsic utility (reflecting its quality, reputation, and timeliness), and storage overhead. This allows caching decisions to simultaneously respond to the real-time popularity, intrinsic value, and cost of the data. Secondly, based on the cache evaluation values, the data is sorted and tiered using a Top-K rule. This automatically stores the most frequently accessed and most valuable "hot" knowledge in a high-speed hot cache, ensuring that LLM inference requests receive extremely low-latency responses. Data with relatively low value and access frequency is stored in a warm or cold cache, achieving a reasonable balance between storage cost and access speed. Finally, the introduced dynamic update mechanism ensures that cached content can be adjusted in real time according to changes in knowledge popularity, utility, and system resource status. When new data is added or space is insufficient, the system can automatically re-evaluate and optimize data distribution, thereby continuously maintaining the overall performance of the cache. This embodiment enables the intelligent and adaptive allocation of limited edge cache resources, ultimately ensuring that LLM can continuously and efficiently acquire the most valuable knowledge.
[0082] Figure 4 A schematic diagram of a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching system is provided for embodiments of the present invention, as shown below. Figure 4 As shown, a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching system includes a data provider module, a validator module, a reputation management module, a data filtering module, a data sharing and notarization module, and a cache management module, wherein: The data provider module is used to listen for knowledge requests on the blockchain network and to store the metadata and storage certificates of the data to be shared on the blockchain. The validator module is used to obtain the data to be shared from the blockchain network, evaluate the data to be shared from multiple preset dimensions, and generate a verification score; The reputation management module, deployed in the blockchain network, is used to update the reputation of data items and the reputation of validator nodes based on the verification score and through a preset reputation calculation model. The data filtering module is used to filter out a set of data items that meet the conditions from all the data to be shared, based on preset constraints, the updated reputation of data items, the multi-dimensional quality assessment results of data items, and a trade-off function composed of data latency and credibility. The data sharing and evidence storage module is used to transfer the selected set of data items from the data provider module to the cache module, and to store the summary of the selection results on the blockchain for evidence storage. The cache management module is used to perform hierarchical storage and dynamic management of the received data item set for use by the edge large language model.
[0083] It is understood that the blockchain-based edge LLM trusted knowledge filtering, sharing and caching system provided by the present invention corresponds to the blockchain-based edge LLM trusted knowledge filtering, sharing and caching methods provided in the foregoing embodiments. The relevant technical features of the blockchain-based edge LLM trusted knowledge filtering, sharing and caching system can be referred to the relevant technical features of the blockchain-based edge LLM trusted knowledge filtering, sharing and caching methods, and will not be repeated here.
[0084] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 5 As shown, an embodiment of the present invention provides an electronic device, including a memory 510, a processor 520, and a computer program 511 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 511, it performs the following steps: S1, the data provider node listens for knowledge requests on the blockchain and uploads the metadata of the data to be shared and the storage certificate to the blockchain for evidence storage. S2, the validator node obtains the data to be shared from the blockchain and evaluates it from multiple preset dimensions to generate a verification score; S3, based on the generated verification score, the blockchain network updates the reputation of the data items provided by the data provider node and the reputation of the verifier node itself through a preset reputation calculation model; S4. Based on the preset constraints, combined with the updated data item reputation, the multi-dimensional quality assessment results of the data items, and the trade-off function composed of data latency and credibility, a set of data items that meet the preset constraints is selected from all the data to be shared. S5, the selected set of data items is transmitted from the corresponding data provider node to the cache node for use by the large language model, and the result summary of this screening is stored on the blockchain for evidence. S6, the cache node performs hierarchical storage and dynamic management of the received set of data items.
[0085] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by the present invention. (See diagram below.) Figure 6 As shown, this embodiment provides a computer-readable storage medium 600, on which a computer program 511 is stored. When the computer program 511 is executed by a processor, it performs the following steps: S1, the data provider node listens for knowledge requests on the blockchain and uploads the metadata of the data to be shared and the storage certificate to the blockchain for evidence storage. S2, the validator node obtains the data to be shared from the blockchain and evaluates it from multiple preset dimensions to generate a verification score; S3, based on the generated verification score, the blockchain network updates the reputation of the data items provided by the data provider node and the reputation of the verifier node itself through a preset reputation calculation model; S4. Based on the preset constraints, combined with the updated data item reputation, the multi-dimensional quality assessment results of the data items, and the trade-off function composed of data latency and credibility, a set of data items that meet the preset constraints is selected from all the data to be shared. S5, the selected set of data items is transmitted from the corresponding data provider node to the cache node for use by the large language model, and the result summary of this screening is stored on the blockchain for evidence. S6, the cache node performs hierarchical storage and dynamic management of the received set of data items.
[0086] This invention provides a blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method, system, and storage medium, constructing a full-process trusted sharing system from knowledge publication, verification, filtering to caching applications. The invention begins with the data provider storing the metadata of the data to be shared on the blockchain. Then, verifier nodes evaluate the data from multiple dimensions and generate verification scores. Based on these scores, the system uses a sophisticated bidirectional reputation calculation model to synchronously update the reputation of both the data item itself and the verifier nodes, thereby incentivizing high-quality data contributions and impartial verification. During the filtering phase, the system comprehensively evaluates the updated reputation, static quality, statistical quality, and task relevance of the data item from multiple dimensions. It also introduces a trade-off function to quantify the trade-off between data latency and credibility. Under multiple preset constraints such as latency, credibility, and adoption quantity, the system intelligently filters out the optimal data set with the goal of maximizing overall utility. Finally, the filtered data is transmitted to caching nodes and stored and dynamically managed hierarchically based on their dynamically calculated caching evaluation values for efficient use by edge large language models.
[0087] This invention solves the core problems of inconsistent data quality, difficulty in constraining the behavior of participants, and difficulty in balancing real-time performance and reliability in edge LLM knowledge sharing, and ultimately achieves synergistic optimization of knowledge updates in three dimensions: trustworthiness, quality, and efficiency.
[0088] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0093] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0094] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A blockchain-based edge LLM trusted knowledge filtering, sharing and caching method, characterized in that, include: S1, the data provider node listens for knowledge requests on the blockchain and uploads the metadata of the data to be shared and the storage certificate to the blockchain for evidence storage. S2, the validator node obtains the data to be shared from the blockchain and evaluates it from multiple preset dimensions to generate a verification score; S3, based on the generated verification score, the blockchain network updates the reputation of the data items provided by the data provider node and the reputation of the verifier node itself through a preset reputation calculation model; S4. Based on the preset constraints, combined with the updated data item reputation, the multi-dimensional quality assessment results of the data items, and the trade-off function composed of data latency and credibility, a set of data items that meet the preset constraints is selected from all the data to be shared. S5, the selected set of data items is transmitted from the corresponding data provider node to the cache node for use by the large language model, and the result summary of this screening is stored on the blockchain for evidence. S6, the cache node performs hierarchical storage and dynamic management of the received set of data items.
2. The blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 1, characterized in that, In step S3, the reputation of the data item Calculated using the following formula: in, To update the weights (0 < <1), This is a credibility penalty coefficient. To verify the weighted mean of the scores, For data credibility, The maximum confidence threshold; The weighted average of the verification scores The calculation model is as follows: in, For data items p The weighted average of the verification scores at time t. For the i-th validator, the data item p of t Verify scores at all times. For the i-th verifier t Reputation at all times For validator nodes, V This is the set of validator nodes.
3. The blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 2, characterized in that, The credibility of the data For all validator nodes, pair data items p The standard deviation of the verification scores is calculated as follows: in, The number of validators.
4. The blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 1, characterized in that, In step S3, the reputation of the validator node Calculated using the following model: in, For the i-th verifier Reputation at all times For validator nodes, For validator nodes Scoring consistency bias, It is used to measure the consistency of its score with the scores of other validator nodes.
5. A blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 1, characterized in that, In step S4, the multi-dimensional quality assessment results of the data items It is a weighted sum of static quality indicators, statistical quality indicators, and task relevance indicators; The static quality index is calculated based on the perplexity of the data item; the statistical quality index is a weighted sum of semantic conflict and redundancy; and the task relevance index is obtained by calculating the matching score between the data item and the current query.
6. A blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 1 or 5, characterized in that, In step S4, the trade-off function consisting of data latency and reliability... , represented as: in, and These are the delay weight and the credibility weight, respectively. The preset maximum allowable delay, For data items p Total delay, This includes verification delays, blockchain evidence storage delays, and data transmission delays.
7. A blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 6, characterized in that, In step S4, the preset constraints include at least one of the following: binary constraints for accepting or not accepting data items, delay constraints, data credibility constraints, and constraints on the maximum number of data items to accept.
8. A blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 7, characterized in that, In step S4, a set of data items that meet preset constraints is selected from all the data to be shared. The optimization objective of the data selection is to maximize the total utility of the selected set of data items, where the utility of a single data item is... This is the sum of the reputation, multi-dimensional quality assessment results, and the compromise function value for the data item.
9. A blockchain-based edge LLM trusted knowledge filtering, sharing, and caching method according to claim 8, characterized in that, Step S6 includes: The time decay access frequency, data item utility and storage overhead corresponding to each data item in the data item set are weighted and combined to obtain the cache evaluation value corresponding to each data item. Based on the cache evaluation value, all data items in the data item set are sorted, and according to the preset Top-K rule, the top K1 data items with the highest cache evaluation value are stored in the hot cache, the next K2 data items are stored in the warm cache, and the remaining data items are stored in the cold cache. When new data items need to be stored in the cache or the cache space is insufficient, the cache evaluation value of the relevant data items is recalculated, and the data item sorting and storage level are dynamically adjusted based on the updated cache evaluation value.
10. A blockchain-based edge LLM trusted knowledge filtering, sharing, and caching system, characterized in that, include: The data provider module is used to listen for knowledge requests on the blockchain network and to store the metadata and storage certificates of the data to be shared on the blockchain. The validator module is used to obtain the data to be shared from the blockchain network, evaluate the data to be shared from multiple preset dimensions, and generate a verification score; The reputation management module, deployed in the blockchain network, is used to update the reputation of data items and the reputation of validator nodes based on the verification score and through a preset reputation calculation model. The data filtering module is used to filter out a set of data items that meet the conditions from all the data to be shared, based on preset constraints, the updated reputation of data items, the multi-dimensional quality assessment results of data items, and a trade-off function composed of data latency and credibility. The data sharing and evidence storage module is used to transfer the selected set of data items from the data provider module to the cache module, and to store the summary of the selection results on the blockchain for evidence storage. The cache management module is used to perform hierarchical storage and dynamic management of the received data item set.