Cross-institution nursing data security sharing platform based on block chain and federal learning

Abstract

The invention relates to the technical field of health state monitoring and personalized intervention, in particular to a cross-institution nursing data security sharing platform based on block chain and federated learning, which comprises a parameter acquisition unit used for monitoring and collecting network average round-trip delay, effective network bandwidth and local data set scale of each participant node, the average transaction confirmation time of the block chain network is obtained; the time penalty quantification unit is used for resolving an effective synchronization time penalty index; the asynchronous anchoring decision unit is used for generating a conventional anchoring signal; the periodic abstract generation unit is used for aggregating the key information in the anchoring period into a periodic abstract and uploading the periodic abstract to the block chain; the self-adaptive optimization unit is used for deciding the number of basic turns of the anchoring cycle to perform feedback adjustment; according to the method, the inherent wooden barrel effect and block chain consensus overhead of federated learning are disclosed, a fuzzy system bottleneck problem is converted into an accurate and quantifiable time cost index, and a scientific basis is provided for subsequent dynamic regulation and control.

Description

Technical Field

[0001] This invention relates to the field of health status monitoring and personalized intervention technology, specifically a cross-institutional nursing data secure sharing platform based on blockchain and federated learning. Background Technology

[0002] In applications that integrate federated learning and blockchain technology to achieve secure cross-institutional data sharing, ensuring the auditability of the model training process is crucial. To achieve this, the computation process of federated learning is typically coupled synchronously with the consensus process of the blockchain, i.e., key training information is recorded on-chain in real time. However, in real-world distributed networks, the network bandwidth, latency, and local data scale of participating nodes exhibit significant heterogeneity, while the transaction confirmation time of the blockchain network is also dynamically changing. Most existing technologies use fixed synchronization frequencies or rely on manual configuration to coordinate off-chain computation and on-chain consensus. This static coupling method lacks the ability to perceive environmental heterogeneity. When some nodes have poor network conditions or on-chain transactions are congested, strict synchronization mechanisms force the entire system to wait for the slowest link, causing the synchronization bottleneck to be amplified dramatically, severely restricting the overall training efficiency of federated learning. Existing methods have failed to establish a quantitative model that can comprehensively evaluate the time cost caused by multi-source heterogeneous factors, thus failing to provide a precise basis for dynamically adjusting synchronization strategies and making it difficult to achieve an effective balance between training efficiency and process auditability.

[0003] Therefore, how to construct a closed-loop adaptive control mechanism that can accurately quantify system synchronization bottlenecks and dynamically decouple computation and consensus processes is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention discloses a cross-institutional nursing data secure sharing platform based on blockchain and federated learning. Specifically, the technical solution of this invention includes:

[0005] The parameter acquisition unit is used to monitor and collect the average round-trip latency, effective network bandwidth, and local dataset size of each participating node, and to obtain the average transaction confirmation time of the blockchain network.

[0006] The time penalty quantification unit is used to calculate the communication time cost of each participating node based on the average round-trip delay of the network, the effective network bandwidth and the size of the local dataset collected by the parameter acquisition unit, and to calculate the effective synchronization time penalty index in combination with the average transaction confirmation time.

[0007] The asynchronous anchoring decision unit is used to compare and analyze the effective synchronization time penalty index with the preset time penalty threshold; when the effective synchronization time penalty index is greater than the preset time penalty threshold, a frequency reduction anchoring signal is generated; when the effective synchronization time penalty index is not greater than the preset time penalty threshold, a normal anchoring signal is generated.

[0008] The periodic summary generation unit is used to respond to the down-frequency anchoring signal or the regular anchoring signal, perform off-chain model training within the corresponding anchoring period, and aggregate the key information within the anchoring period into a periodic summary for uploading to the blockchain;

[0009] The adaptive optimization unit monitors the actual total time of the anchoring cycle and calculates the ideal time of the anchoring cycle. Based on the ratio of the actual total time to the ideal time, it provides feedback adjustment to the number of base rounds used to determine the anchoring cycle.

[0010] Preferably, the calculation process for communication time cost is as follows:

[0011] The quotient is obtained by dividing the local dataset size by the effective network bandwidth, and then summing the quotient with the average round-trip delay of the network. The resulting value is set as the communication time cost.

[0012] The preferred method for calculating the effective synchronization time penalty index is as follows:

[0013] Obtain the maximum communication time cost among all participating nodes and weight it according to data heterogeneity; sum the processed value with the weighted average transaction confirmation time, and set the resulting value as the effective synchronization time penalty index.

[0014] The preferred decision-making process for anchoring the period is as follows:

[0015] Divide the effective synchronization time penalty index by the preset time penalty threshold, multiply the resulting quotient by the number of base rounds, and round up the product. The resulting value is set as the anchoring period.

[0016] Preferably, the period summary generation unit is used to concatenate the hash values ​​of the aggregation model parameters at the end of the period, the hash values ​​of the contribution proofs of each participant in the period, and the start and end timestamps of the period into a byte string, and perform hash function processing on the concatenated byte string to generate a period summary.

[0017] The preferred, ideal time-consuming calculation process is as follows:

[0018] Multiply the number of rounds included in the anchor period of the current period by the maximum communication time cost among all participating nodes, sum the product with the average transaction confirmation time, and set the resulting value as the ideal time consumption.

[0019] Preferably, the feedback adjustment process for the number of base rounds is as follows:

[0020] Divide the actual total time by the ideal time, apply damped weighting to the quotient, multiply the result by the number of base rounds in the current cycle, and set the product as the number of base rounds in the next cycle.

[0021] Preferably, it also includes a hierarchical governance unit;

[0022] The hierarchical governance unit is used to obtain the preset proposal type impact factor, the reputation score of the proposal initiator, and the estimated amount of the proposal's potential impact on the overall situation; and, in combination with the proposal type impact factor, the reputation score of the proposal initiator, and the estimated amount of the proposal's potential impact on the overall situation, to calculate the governance importance score.

[0023] Preferably, the hierarchical governance unit is also used to: trigger the on-chain voting process when the governance importance score exceeds a preset governance threshold; and not trigger the on-chain voting process when the governance importance score does not exceed the preset governance threshold.

[0024] Preferably, the reputation score of the proposal initiator is calculated based on the validity of the initiator's contribution proof in the historical anchoring period and the stability of node parameters reflected by communication time cost.

[0025] The cross-institutional nursing data secure sharing platform based on blockchain and federated learning provided by this invention aims to overcome the synchronization bottleneck amplification effect caused by the strong coupling of federated learning and blockchain in heterogeneous environments in existing technologies. Compared with existing technologies, this invention has the following beneficial effects:

[0026] 1. This invention establishes a mechanism for accurately quantifying system performance bottlenecks. Through a parameter acquisition unit, it monitors in real time the network latency, effective bandwidth, local data scale, and transaction confirmation time of each participating node, as well as the blockchain network. A time penalty quantification unit then models these heterogeneous parameters to calculate an effective synchronization time penalty index that comprehensively reflects communication time costs and the impact of data heterogeneity. This index profoundly reveals the inherent bottleneck effect of federated learning and the consensus overhead of blockchain, transforming the fuzzy system bottleneck problem into a precise and quantifiable time cost indicator. This provides a scientific basis for subsequent dynamic control, far exceeding the isolated and one-sided assessments of performance bottlenecks in existing technologies.

[0027] 2. This invention constructs an asynchronous anchoring decision and execution system for flexible decoupled computation and consensus. The asynchronous anchoring decision unit compares the aforementioned quantified time penalty index with a preset time penalty threshold to dynamically determine the anchoring signal that is suitable for the current system state. The periodic summary generation unit responds to this signal, performs multiple rounds of off-chain model training within an anchoring period, and aggregates key information such as the hash of aggregated model parameters, hash of each party's contribution proof, and timestamp into a single periodic summary and uploads it to the blockchain. This design breaks the constraint that each round of computation in the traditional scheme must be rigidly synchronized with the blockchain, realizes dynamic scaling of on-chain evidence storage granularity, and greatly reduces the interaction frequency with the blockchain while ensuring that the process is auditable and tamper-proof, thus significantly improving the overall efficiency of model training.

[0028] 3. This invention introduces a closed-loop adaptive optimization feedback system; the adaptive optimization unit monitors the actual total time of each anchoring cycle and compares it with the ideal time calculated based on the current system parameters to quantify the performance deviation; based on this deviation, the system can continuously adjust the number of basic rounds that determine the length of the anchoring cycle; this mechanism gives the platform the ability to self-calibrate and evolve, enabling it to automatically adapt to the dynamic changes in the network environment and data distribution, and continuously maintain itself at the best balance between model training efficiency and process auditability, solving the rigidity and lag problems caused by the use of fixed parameters or reliance on manual tuning in existing technologies;

[0029] 4. This invention designs an efficient and secure hierarchical governance architecture. The hierarchical governance unit calculates a governance importance score by comprehensively evaluating the impact factors of proposal types, the reputation score of the proposal initiator, and the potential impact of the proposal. This scoring system innovatively links a node's historical technical performance, such as the validity of contribution proofs and communication stability, to its voice in community governance. Based on this score, the system can automatically divert proposals of different importance, triggering highly secure on-chain voting for matters with significant impact, while delegating routine matters to efficient off-chain processes. This not only avoids all governance decisions being congested in inefficient on-chain voting channels but also constructs a positive incentive mechanism, encouraging all participants to contribute stable and high-quality resources to the platform, thereby ensuring the long-term health and positive development of the entire data sharing ecosystem from a higher dimension. Attached Figure Description

[0030] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0031] Figure 1 This is a structural diagram of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0033] Example 1:

[0034] Please see Figure 1 A cross-institutional nursing data security sharing platform based on blockchain and federated learning includes:

[0035] The parameter acquisition unit is used to monitor and collect the average round-trip latency, effective network bandwidth, and local dataset size of each participating node, and to obtain the average transaction confirmation time of the blockchain network.

[0036] The time penalty quantification unit is used to calculate the communication time cost of each participating node based on the average round-trip delay of the network, the effective network bandwidth and the size of the local dataset collected by the parameter acquisition unit, and to calculate the effective synchronization time penalty index in combination with the average transaction confirmation time.

[0037] The asynchronous anchoring decision unit is used to compare and analyze the effective synchronization time penalty index with the preset time penalty threshold; when the effective synchronization time penalty index is greater than the preset time penalty threshold, a frequency reduction anchoring signal is generated; when the effective synchronization time penalty index is not greater than the preset time penalty threshold, a normal anchoring signal is generated.

[0038] The periodic summary generation unit is used to respond to the down-frequency anchoring signal or the regular anchoring signal, perform off-chain model training within the corresponding anchoring period, and aggregate the key information within the anchoring period into a periodic summary for uploading to the blockchain;

[0039] The adaptive optimization unit is used to monitor the actual total time of the anchoring cycle and calculate the ideal time of the anchoring cycle. Based on the ratio of the actual total time to the ideal time, it provides feedback adjustment to the number of base rounds used to determine the anchoring cycle.

[0040] This invention provides a cross-institutional nursing data secure sharing platform based on blockchain and federated learning, comprising: a parameter acquisition unit, a time penalty quantification unit, an asynchronous anchoring decision unit, a periodic summary generation unit, and an adaptive optimization unit;

[0041] This embodiment aims to address the technical pain point of the synchronization bottleneck amplification effect caused by the strong coupling between federated learning and blockchain in heterogeneous network and data environments in the existing technology; the platform achieves both model training efficiency and process auditability through a closed-loop adaptive system that dynamically decouples the computation and consensus process.

[0042] The parameter acquisition unit aims to provide real-time, multi-dimensional data input for subsequent performance bottleneck quantification and dynamic strategy adjustment. In this embodiment, this unit is implemented through agent programs deployed on the clients of various participating parties, such as the data servers of various nursing institutions. Specifically, this unit is responsible for monitoring and collecting key parameters affecting synchronization performance, and its operation is as follows:

[0043] Average round-trip latency of the network With effective network bandwidth The measurement is obtained by periodically sending standard-sized probe packets to the federated learning aggregator through a client agent and recording the response time. This refers to the average time it takes for a data packet to travel from the client to the aggregator and back, measured in seconds. Its function is to measure the latency of network communication; This refers to the actual data transfer rate available within a specific time period, measured in megabytes per second. Its function is to measure the throughput of network communication;

[0044] Size of local dataset The size or number of records of the nursing dataset file read directly from the local storage by the local client agent, in megabytes. Its function is to quantify the load of data transmission;

[0045] Average transaction confirmation time This is obtained by actively querying the API interface of the connected blockchain network, reflecting the current congestion status and consensus efficiency of the blockchain network, in seconds. Its function is to quantify the time cost of on-chain evidence storage;

[0046] The purpose of the time penalty quantification unit is to unify multiple heterogeneous performance influencing factors into a single, predictable time cost index with clear physical meaning. In this embodiment, based on data collected by the parameter acquisition unit, this unit constructs a mathematical model called the effective synchronization time penalty index Ψ. To clarify its internal logic, this unit first calculates the communication time cost of each participating node n. The calculation method is as follows:

[0047]

[0048] in, The estimated total communication time, in seconds, required for node n to complete one federated learning data exchange, such as uploading model updates. Its source is from the preceding sequence , , The parameters are calculated; this formula integrates network latency and data transmission time, comprehensively reflecting the communication performance of the node;

[0049] Based on the aforementioned communication time costs, this unit, considering the worst-case scenario among all participants and the overhead of the blockchain, calculates the effective synchronization time penalty index Ψ:

[0050]

[0051] Where Ψ represents the expected total time overhead for the system to complete a fully synchronized federated learning interaction, including off-chain computation and on-chain consensus, in seconds. Its source is derived from comprehensive calculations within this unit; This reflects the "barrel effect" in federated learning, where the slowest node must be waited for. It is data heterogeneity, which is a dimensionless pure number obtained by calculating the KL divergence between the local data distribution and the global data distribution. Its source is the historical value obtained by the federated learning framework after the training of the previous anchoring cycle. Its role is to quantify the impact of data distribution differences on the model convergence speed. The weights are dimensionless, and their calibration process is crucial to ensuring the accuracy of the model. To clarify the calibration method, it is necessary to clearly distinguish between the variables used in the calibration process and those used during model runtime. Calibration is performed in a benchmark environment containing a set of controllable test nodes. During the calibration phase, multiple different combinations of network parameters and dataset sizes are set for each test node. And record the total time actually measured to complete one round of synchronous interaction. Therefore, a set of independent variables can be obtained. With dependent variable The data points constituted This represents the average round-trip latency of the network under benchmark testing conditions. Represents the effective network bandwidth under benchmark testing conditions. This represents the size of the local dataset in the benchmark testing environment; by performing multiple linear regression fitting on these data points, the solution is obtained. and The best estimate; this index transforms the fuzzy performance bottleneck problem into a quantifiable time cost, providing a precise basis for subsequent asynchronous decisions;

[0052] The asynchronous anchoring decision unit aims to dynamically adjust the granularity of blockchain notarization based on the system's current estimated time cost, breaking the rigid synchronization relationship between computation and consensus. In this embodiment, the unit combines the effective synchronization time penalty index Ψ calculated by the time penalty quantization unit with a preset time penalty threshold. Perform comparative analysis; This refers to the maximum acceptable single-round synchronization time cost of the system, such as 10 seconds. Its purpose is to serve as a benchmark for triggering asynchronous strategies. It is derived from empirical values ​​determined based on the business's requirements for real-time performance and stress testing of system performance.

[0053] when When this happens, it means that if the current system adopts synchronous mode, an unacceptable time delay will occur. At this time, the unit generates a frequency-reduced anchoring signal.

[0054] when When this occurs, it means that the system performance is still within an acceptable range, and the unit generates a regular anchoring signal.

[0055] The cycle summary generation unit aims to perform specific off-chain computations and reliably solidify key process information within a computation cycle in a lightweight manner. In this embodiment, in response to a signal generated by the asynchronous anchoring decision unit, this unit performs N rounds of federated learning model training and aggregation within a time window called the anchoring cycle. This process is completed entirely off-chain and is unaffected by blockchain consensus latency. The anchoring cycle refers to a set of consecutive federated learning communication rounds, after which it interacts with the blockchain once. Its function is to bundle multiple computations and amortize the cost of a single on-chain interaction. At the end of the cycle, the unit will generate all key information within the cycle, such as the hash of the global model parameters at the end of the cycle. Hash of each participant's contribution proof Start and end timestamps of the period Through collision-resistant hash functions, such as SHA-256, a single periodic digest is aggregated. This summary is uploaded to the blockchain as the transaction content, while the detailed log is stored in off-chain distributed storage.

[0056] The adaptive optimization unit aims to construct a top-level closed-loop feedback mechanism, enabling the system to self-adjust core control parameters based on actual operating performance, thereby dynamically seeking optimization between efficiency and auditability. In this embodiment, after each anchoring cycle, the unit first monitors and records the actual total time consumed in that cycle. At the same time, it will be based on the actual communication time cost of each node within this cycle. Confirmation time with blockchain Calculate the ideal time for one anchoring cycle. This serves as a benchmark for performance evaluation; finally, the unit is evaluated based on the actual total time consumed. Compared to the ideal time The ratio of the base rounds used to determine the next anchoring cycle. Perform feedback adjustments;

[0057] This invention constructs a complete closed-loop adaptive system through the collaborative work of the aforementioned units. First, it accurately predicts the cost Ψ of the synchronization bottleneck by collecting parameters and quantifying time penalties. Then, based on this cost, it dynamically determines the degree of decoupling between computation and consensus, and significantly reduces the frequency of on-chain interactions while ensuring traceability through periodic summaries. Most importantly, the adaptive optimization unit enables the system to learn from historical performance and continuously self-calibrate its core parameters. Thus, in complex and ever-changing network and data environments, it always maintains the best balance between efficient model training and reliable process recording, solving the problem of the synchronous bottleneck amplification effect of existing technologies.

[0058] Example 2:

[0059] The calculation process for communication time cost is as follows:

[0060] Divide the local dataset size by the effective network bandwidth to obtain the quotient, and sum the quotient with the average round-trip delay of the network. The resulting value is set as the communication time cost.

[0061] Based on the platform described in Example 1, the calculation process of communication time cost is as follows: the local dataset size is divided by the effective network bandwidth to obtain the quotient, and the quotient is summed with the average round-trip delay of the network. The resulting value is set as the communication time cost.

[0062] This embodiment quantizes the communication time cost in the time penalty unit. The specifics of the calculation method; as mentioned above, The purpose is to accurately estimate the total time required for a single node to complete one data interaction. The specific calculation formula is as follows:

[0063]

[0064] The calculation logic of this formula stems from the physical decomposition of the network communication process; the total time of a single data exchange mainly consists of two parts: one is the delay time of data packet transmission in the network, i.e., the average round-trip delay of the network. Secondly, there is the transmission time required to transmit the data content itself; this transmission time is determined by the size of the local dataset. (Total data volume) and effective network bandwidth The transmission rate is determined by the ratio of the two. This gives the transmission time; adding these two independent time costs together gives a comprehensive result, measured in seconds. Communication time cost;

[0065] Wherein, all parameters in the formula , , All data are acquired in real time by the parameter acquisition unit through the aforementioned probe measurement and local reading methods, ensuring that the calculation results can truly reflect the current actual state of the node;

[0066] By combining the effects of latency and transmission bandwidth in this calculation method, the present invention can more accurately characterize the communication performance of nodes. Compared with schemes that only consider latency or bandwidth as a single indicator, this scheme provides a more comprehensive and accurate evaluation of node performance. For example, it can distinguish between two different performance bottleneck types: high network latency but large bandwidth and low network latency but small bandwidth. This provides a more reliable data foundation for the accurate construction of the upper-layer effective synchronization time penalty index Ψ, thereby improving the accuracy of dynamic decision-making in the entire system.

[0067] It should be noted that this formula is a first-order approximation model of communication time cost, effectively capturing the two core influencing factors: network latency and transmission throughput. In practical applications, time costs such as local computation overhead and data serialization can be considered as included in the average round-trip delay of the network, which is periodically measured by probes. In this context, cost is considered a secondary concern compared to network transmission time. This model, while maintaining computational simplicity, provides sufficiently accurate cost estimates to support higher-level decision-making.

[0068] Example 3:

[0069] The calculation process for the effective synchronization time penalty index is as follows:

[0070] Obtain the maximum communication time cost among all participating nodes and weight it according to data heterogeneity; sum the processed value with the weighted average transaction confirmation time, and set the resulting value as the effective synchronization time penalty index;

[0071] Based on the platform described in Example 1, the calculation process of the effective synchronization time penalty index is as follows: obtain the maximum communication time cost among all participating nodes and perform weighted processing in combination with data heterogeneity; sum the processed value with the weighted average transaction confirmation time, and set the resulting value as the effective synchronization time penalty index;

[0072] This embodiment is a concretization of the calculation process of the effective synchronization time penalty index Ψ in the time penalty quantification unit; as mentioned above, the purpose of Ψ is to estimate the total time cost required for the system to complete a full round of synchronization interaction, and its specific calculation formula is as follows:

[0073]

[0074] The underlying logic of this formula integrates the "weakest link" effect of federated learning, the impact of data heterogeneity, and the independent overhead of blockchain consensus.

[0075] Quantifying the "barrel effect": In federated learning, each round of aggregation must wait for the slowest participant to complete its model upload. Therefore, the solution process first... The operation is to obtain the largest communication time cost among all participating nodes. This serves as the benchmark for communication time in this round of the chain;

[0076] Impact of data heterogeneity: Data heterogeneity This refers to the degree of inconsistency in the data distribution among the participating parties. Specifically, if we assume that the participating parties... The local data distribution is The global data distribution is as follows Then data heterogeneity can be defined as ,in, This represents the Kullback-Leibler divergence, used to measure the difference between the local and global data distributions. At the end of each anchoring cycle, the system estimates the current global distribution based on the model updates uploaded by each party. and the local distribution known to each participant. Compare, and thus update Value. During the initial anchoring period of the system, It can be initialized to 0, indicating that no prior assumptions are made about the data distribution. If, for some reason, a valid result cannot be successfully calculated after a certain period ends... The system can either reuse the historical value from the previous cycle or reset it to 0 to ensure the robustness of the decision-making process.

[0077] It increases the number of training epochs required for the model to converge, which is simplified here as a weighted effect on the time cost of a single epoch; by multiplying by This term quantifies the negative impact of data distribution differences on training efficiency as an increase in time cost; to ensure the uniqueness of the computation, this... Specifically refers to The data heterogeneity index of the node with the highest determined communication time cost.

[0078] The linear weighting here is an engineering simplification that maps the convergence problem at the algorithm level to the time cost at the system level. Its purpose is not to accurately predict the total number of rounds required for convergence, but rather to assign higher risk weights to nodes with poorer data distribution at the granularity of a single round of decision-making, thereby reflecting the additional training overhead that may result from data heterogeneity in the time penalty exponent. This simplification is effective for macro-level control.

[0079] Multi-source cost fusion: This involves weighting communication penalties based on data heterogeneity. With weight Adjusted average transaction confirmation time Summing these two time costs sequentially yields the estimated total time cost Ψ.

[0080] This solution process does not simply add up the various time parameters, but rather deeply understands and models the core bottleneck sources in distributed collaborative systems; by adopting Instead of the average, it accurately captures the barrel effect; by introducing This establishes a direct mathematical link between the data heterogeneity problem at the algorithm level and the time cost problem at the system level. This makes the Ψ index not only a time estimate, but also a multi-dimensional comprehensive assessment of the overall health of the system. Its accuracy far exceeds that of existing technologies that isolate network latency or transaction time, providing extremely accurate decision input for subsequent asynchronous anchoring strategies.

[0081] Example 4:

[0082] The decision-making process for anchoring the cycle is as follows:

[0083] Divide the effective synchronization time penalty index by the preset time penalty threshold, multiply the resulting quotient by the number of base rounds, and round up the product. The resulting value is set as the anchoring period.

[0084] Based on the platform described in Example 1, the decision process for the anchoring period is as follows: divide the effective synchronization time penalty index by the preset time penalty threshold, multiply the resulting quotient by the number of basic rounds, and round up the product to obtain the anchoring period.

[0085] This embodiment is a concretization of the decision-making process of how the asynchronous anchoring decision unit determines the anchoring period N. The specific decision-making formula is as follows:

[0086]

[0087] This decision-making logic originates from the threshold triggering and proportional adjustment mechanism in control theory, aiming to dynamically adjust the frequency of interaction with the blockchain;

[0088] Through ratio Evaluate the currently estimated effective synchronization time penalty index Ψ relative to the system's acceptable time penalty threshold. Exceeding the limit; if The ratio is no greater than 1; if The ratio is greater than 1, and the larger the value, the more severe the performance bottleneck.

[0089] Ratio this to the number of base rounds Multiplication; number of base rounds This refers to the ideal state of the system, where Ψ is approximately equal to... At that time, the reference anchoring frequency can be set, for example, to 5 rounds; its initial value The anchoring period N can be determined comprehensively based on the business's basic requirements for audit granularity and the performance evaluation results in a benchmark testing environment, serving as the starting point for the system's adaptive adjustment; when the system performance is good, the anchoring period N is taken as... When the system encounters a performance bottleneck, the anchoring period N will increase linearly with the increase of Ψ.

[0090] Use the rounding up function Ensure that the calculated number of rounds N is an executable positive integer;

[0091] This decision-making process establishes a direct and automatic control path from performance bottleneck quantification to system behavior adjustment. When the system time cost Ψ increases, the anchoring period N automatically lengthens, thereby intelligently reducing the frequency of interaction with the blockchain and effectively alleviating performance pressure. Conversely, when performance recovers, N automatically shortens to pursue finer audit granularity. This dynamic and quantitative decision-making mechanism enables the system to be flexible in responding to strong conditions and precise in responding to weak conditions, completely avoiding the rigidity and lag problems caused by fixed anchoring periods or manual parameter adjustments in existing technologies.

[0092] Example 5:

[0093] The cycle summary generation unit is used to concatenate the hash values ​​of the aggregate model parameters at the end of the cycle, the hash values ​​of the contribution proofs of each participant in the cycle, and the start and end timestamps of the cycle into a byte string, and perform hash function processing on the concatenated byte string to generate a cycle summary.

[0094] Based on the platform described in Example 1, the period summary generation unit is used to concatenate the hash values ​​of the aggregated model parameters at the end of the period, the hash values ​​of the contribution proofs of each participant in the period, and the start and end timestamps of the period into a byte string, and perform hash function processing on the concatenated byte string to generate a period summary.

[0095] This embodiment describes how the periodic summary generation unit generates periodic summaries. The process aims to compress all key information within an anchoring period into a fixed-length, tamper-proof, lightweight data set for uploading to the blockchain. The specific generation formula is as follows:

[0096]

[0097] This generation logic draws on the Merkle tree concept from cryptography, ensuring the integrity and verifiability of the data;

[0098] in, This is a standard collision-resistant hash function, such as SHA-256, which can transform an input of arbitrary length into a unique digest of fixed length.

[0099] : It refers to the hash value of the parameter file of the final aggregated global model at the end of the anchoring period; its function is to solidify the final calculation results of this period.

[0100] : This refers to the hash value or other contribution proof of participant m's contribution in each round of the cycle, updated locally; each participant's contribution proof is hashed first and then incorporated into the overall digest; its purpose is to ensure that each participant's contribution is traceable and undeniable.

[0101] : This refers to the start and end timestamps of the anchoring period; its purpose is to provide a clear temporal context for the summary.

[0102] This symbol represents a sequential concatenation operation of byte strings; the unit concatenates all the above data items into a long byte string in a predetermined order;

[0103] Perform a hash function on the entire concatenated long string. The final result is the periodic summary. ;

[0104] The periodic summary generated in this way possesses extremely high information compression rate and strong security features; it only requires a tiny amount of space in a single blockchain transaction, effectively solidifying all the key information from N rounds of training within a cycle; any tampering with the detailed logs stored off-chain, such as a model update from a participant in a particular round, will lead to a recalculated periodic summary. Inconsistencies with values ​​recorded on the chain can be easily detected. This model of on-chain evidence storage and off-chain details reduces the amount of data on the chain by N orders of magnitude without sacrificing final auditability. It is one of the core technical means to decouple computation and consensus and significantly improve system operating efficiency.

[0105] Example 6:

[0106] The calculation process for the ideal time consumption is as follows:

[0107] Multiply the number of rounds included in the anchor period of the current period by the maximum communication time cost among all participating nodes, sum the product with the average transaction confirmation time, and set the resulting value as the ideal time consumption.

[0108] Based on the platform described in Example 1, the calculation process for the ideal time consumption is as follows: multiply the number of rounds included in the anchoring period of the current period by the maximum communication time cost among all participating nodes, sum the product with the average transaction confirmation time, and set the resulting value as the ideal time consumption.

[0109] This embodiment describes the ideal time consumption in the adaptive optimization unit. The concretization of the calculation process; The purpose is to provide a reasonable performance benchmark based on the current system state for the actual performance of the current k-th anchoring period. The specific calculation formula is as follows:

[0110]

[0111] This formula is designed to simulate the time required to complete the current anchoring cycle under ideal conditions;

[0112] pass Find the largest communication time cost among all participating nodes; this is consistent with the system's understanding of the "barrel effect" as a performance bottleneck, representing the expected time in reality when a round of synchronization must wait for the slowest node.

[0113] The maximum time consumption is compared with the actual number of rounds in the current k-th cycle. Multiply; This refers to the actual number of rounds executed in the current cycle, the value of which is determined by the decision-making process in Example 4 at the end of the previous cycle; product This represents the total time theoretically required to complete all off-chain computations in the current cycle;

[0114] In addition to the average transaction confirmation time required for one on-chain operation Because an anchoring cycle only involves one on-chain operation, it is only added once. ;

[0115] The key advantage of this ideal time calculation method lies in its self-consistency and completeness; it does not use any fixed, preset ideal time parameters, but is entirely based on the real-time measurements taken by the parameter acquisition unit, reflecting the current actual system state. and Values ​​are calculated dynamically; this means that performance benchmarks are calculated dynamically. It will automatically adjust according to changes in network conditions and blockchain congestion, thus enabling a fairer and more accurate assessment of the actual total time spent. The quality of the benchmark; the establishment of this dynamic benchmark provides a more reliable input for subsequent basic rounds of feedback adjustment, avoiding erroneous adjustments caused by benchmark rigidity.

[0116] Example 7:

[0117] The feedback adjustment process for the number of base rounds is as follows:

[0118] Divide the actual total time by the ideal time, apply damped weighting to the quotient, multiply the result by the number of base rounds in the current cycle, and set the product as the number of base rounds in the next cycle.

[0119] Based on the platform described in Example 1, the feedback adjustment process for the number of base rounds is as follows: divide the actual total time by the ideal time, perform damped weighting on the quotient, multiply the result by the number of base rounds in the current cycle, and set the product as the number of base rounds in the next cycle.

[0120] This embodiment describes the number of rounds for the core control parameters in the adaptive optimization unit. The feedback adjustment process is concretized; this process is the top-level optimization logic of the system, constituting a closed-loop adaptive adjustment between efficiency and auditability, and its specific adjustment formula is as follows:

[0121]

[0122] This logic originates from the proportional feedback control concept in control theory and is implemented through a multiplicative form, making it more robust in a distributed environment.

[0123] Through ratio Calculate the performance deviation for this cycle; This is the actual time spent. The ratio represents the ideal time taken; if the ratio is greater than 1, it indicates that the system is running slower than expected; if the ratio is less than 1, it indicates that there is performance margin.

[0124] Perform this performance deviation ratio Damped weighted processing of powers of power; It is a dimensionless damping coefficient, typically ranging from 0 to 1, such as 0.5. Its function is to prevent the system from oscillating due to overexcited response. Its value is determined through system simulation experiments, inputting a series of typical step or impulse performance disturbances and observing different... Values The convergence rate and overshoot of the parameters are ultimately selected to enable the system to reach stability as quickly as possible while minimizing oscillations. value;

[0125] Multiply the performance deviation after damping weighting by the number of base rounds in the current cycle. The result obtained This is the new base round number set for the next cycle decision;

[0126] when If the ratio is greater than 1, the system will automatically increase it. This means that under similar Ψ exponents in the future, the calculated anchoring period N will be longer, thus amortizing the time cost by reducing the frequency of on-chain activity; conversely, when At that time, the system will adjust appropriately. When performance is sufficient, a more precise anchoring cycle is pursued.

[0127] The establishment of this feedback adjustment mechanism endows the system of this invention with the ability to self-evolve; it no longer relies on a set of static parameters that require manual intervention for optimization, but can automatically and continuously optimize its core asynchronous anchoring strategy based on real-world performance feedback; this closed-loop adaptive capability ensures that the system can operate stably at the optimal dynamic balance point between efficiency and reliability for a long time, regardless of the external environment, which is a major technological advancement that cannot be matched by existing static or open-loop systems.

[0128] To ensure the robustness of the algorithm, in actual computation, the ideal computation time is... Set a very small positive lower bound. ,For example Seconds, i.e., the denominator actually used is This prevents computational overflow issues caused by the denominator approaching zero in rare, extreme cases, ensuring that the adjustment mechanism behaves as expected under all possible inputs.

[0129] Example 8:

[0130] The platform also includes a hierarchical governance unit;

[0131] The hierarchical governance unit is used to obtain the preset proposal type impact factor, the reputation score of the proposal initiator, and the estimated amount of the proposal's potential impact on the overall situation; and combined with the proposal type impact factor, the reputation score of the proposal initiator, and the estimated amount of the proposal's potential impact on the overall situation, to calculate the governance importance score.

[0132] The tiered governance unit is also used to: trigger an on-chain voting process when the governance importance score exceeds a preset governance threshold; and not trigger an on-chain voting process when the governance importance score does not exceed the preset governance threshold.

[0133] The proposal initiator's reputation score is calculated based on the validity of the initiator's contribution proofs during the historical anchoring period, as well as the stability of node parameters reflected by communication time costs.

[0134] Based on the platform described in Example 1, a hierarchical governance unit is also included. This hierarchical governance unit is used to obtain a preset proposal type impact factor, the reputation score of the proposal initiator, and the estimated global potential impact of the proposal. It then calculates a governance importance score by combining the proposal type impact factor, the reputation score of the proposal initiator, and the estimated global potential impact of the proposal. The hierarchical governance unit is also used to trigger an on-chain voting process when the governance importance score exceeds a preset governance threshold. The reputation score of the proposal initiator is calculated based on the validity of the initiator's contribution proofs during the historical anchoring period and the stability of node parameters reflected by communication time costs.

[0135] This embodiment is a concretization of the hierarchical governance unit introduced by the platform at the management level; the purpose of this unit is to distinguish between high-frequency operational tasks and low-frequency, high-impact governance decisions, and to design an efficient and secure automated decision-making process for the latter.

[0136] The core of this unit is calculating a governance importance score. Its calculation formula draws on the multi-criteria decision analysis model:

[0137]

[0138] in, The importance score for governance is a normalized dimensionless value, which is derived from the calculation of this unit.

[0139] : Proposal type influence factor, based on the type of governance proposal, such as: new organization joining, global model version release, system parameter modification, etc., the preset influence weight, which is a constant set by the alliance governance committee according to the governance rules during system initialization;

[0140] The reputation score of the proposal initiator is crucial for achieving parameter linkage design between modules; reputation score It is not a static value, but a dynamically calculated value; its calculation is based on: a) the validity of the contribution proofs submitted by the initiator during the historical anchoring period, generated by the period summary generation unit. The evaluation will be based on whether the subsequent verification passes; b) the stability of the initiating node, which is determined by its historical communication time cost. The fluctuations reflect this; a node with a true historical contribution and a stable network, its... It will be higher;

[0141] The estimated potential impact of the proposal on the overall situation is a quantitative assessment provided by the proposal initiator when submitting the proposal, based on a template, and undergoes preliminary manual or automated review.

[0142] The corresponding dimensionless weighting coefficients are set by the Alliance Governance Committee to balance the importance of different factors.

[0143] Calculate After scoring, the tiered governance unit will compare it with a preset governance threshold. The threshold is set based on governance risk control strategies. For example, the governance committee could set a target requiring that proposals with the highest estimated impact, or those initiated by low-reputation nodes, accounting for the top 10% of the total proposals, be forcibly included in on-chain voting. Based on this, statistical analysis of historical proposal data can be used to calculate the threshold. The specific value;

[0144] when When this occurs, it indicates that the proposal is of great importance, and the unit will automatically trigger the on-chain smart contract to initiate a formal voting process that requires confirmation from multiple parties;

[0145] when If the proposal has a minor impact or is a routine operation, it will be handled through a lightweight off-chain consensus mechanism or directly approved by an administrator with a specific role, without the need to initiate a costly on-chain vote.

[0146] The introduction of the hierarchical governance model greatly improves the efficiency and security of the system at the management level. Through a scoring mechanism, it diverts transactions of different natures to the most efficient processing path, avoiding all decisions being crowded into inefficient on-chain voting channels. The most innovative aspect is that it directly links the historical technical performance of nodes with their governance authority. This design not only makes governance decisions more fair and credible, but also, from an economic and game theory perspective, positively incentivizes all participants to actively and stably contribute high-quality data and computing resources to federated learning, thereby ensuring the long-term healthy development of the entire platform and the virtuous cycle of the data sharing ecosystem at a higher level.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cross-institutional nursing data secure sharing platform based on blockchain and federated learning, characterized in that: include: The parameter acquisition unit is used to monitor and collect the average round-trip latency, effective network bandwidth, and local dataset size of each participating node, and to obtain the average transaction confirmation time of the blockchain network. The time penalty quantification unit is used to calculate the communication time cost of each participating node based on the average round-trip delay of the network, the effective network bandwidth and the size of the local dataset collected by the parameter acquisition unit, and to calculate the effective synchronization time penalty index in combination with the average transaction confirmation time. The asynchronous anchoring decision unit is used to compare and analyze the effective synchronization time penalty index with the preset time penalty threshold; when the effective synchronization time penalty index is greater than the preset time penalty threshold, a frequency reduction anchoring signal is generated; when the effective synchronization time penalty index is not greater than the preset time penalty threshold, a normal anchoring signal is generated. The periodic summary generation unit is used to respond to the down-frequency anchoring signal or the regular anchoring signal, perform off-chain model training within the corresponding anchoring period, and aggregate the key information within the anchoring period into a periodic summary for uploading to the blockchain; The adaptive optimization unit monitors the actual total time of the anchoring cycle and calculates the ideal time of the anchoring cycle. Based on the ratio of the actual total time to the ideal time, it provides feedback adjustment to the number of base rounds used to determine the anchoring cycle.

2. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning as described in claim 1, characterized in that, The calculation process for communication time cost is as follows: The quotient is obtained by dividing the local dataset size by the effective network bandwidth, and then summing the quotient with the average round-trip delay of the network. The resulting value is set as the communication time cost.

3. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 1, characterized in that, The calculation process for the effective synchronization time penalty index is as follows: Obtain the maximum communication time cost among all participating nodes and weight it according to data heterogeneity; sum the processed value with the weighted average transaction confirmation time, and set the resulting value as the effective synchronization time penalty index.

4. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 1, characterized in that, The decision-making process for anchoring the cycle is as follows: Divide the effective synchronization time penalty index by the preset time penalty threshold, multiply the resulting quotient by the number of base rounds, and round up the product. The resulting value is set as the anchoring period.

5. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 1, characterized in that, The cycle summary generation unit is used to concatenate the hash values ​​of the aggregated model parameters at the end of the cycle, the hash values ​​of the contribution proofs of each participant in the cycle, and the start and end timestamps of the cycle into a byte string, and then perform hash function processing on the concatenated byte string to generate a cycle summary.

6. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 1, characterized in that, The calculation process for the ideal time consumption is as follows: Multiply the number of rounds included in the anchor period of the current cycle by the maximum communication time cost among all participating nodes, sum the product with the average transaction confirmation time, and set the resulting value as the ideal time consumption.

7. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 1, characterized in that, The feedback adjustment process for the number of base rounds is as follows: Divide the actual total time by the ideal time, apply damped weighting to the quotient, multiply the result by the number of base rounds in the current cycle, and set the product as the number of base rounds in the next cycle.

8. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 1, characterized in that, It also includes a hierarchical governance unit; The hierarchical governance unit is used to obtain the preset proposal type impact factor, the reputation score of the proposal initiator, and the estimated amount of the proposal's potential impact on the overall situation; and, in combination with the proposal type impact factor, the reputation score of the proposal initiator, and the estimated amount of the proposal's potential impact on the overall situation, to calculate the governance importance score.

9. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 8, characterized in that, The tiered governance unit is also used to: trigger the on-chain voting process when the governance importance score exceeds the preset governance threshold; and not trigger the on-chain voting process when the governance importance score does not exceed the preset governance threshold.

10. The cross-institutional nursing data secure sharing platform based on blockchain and federated learning according to claim 8, characterized in that, The proposal initiator's reputation score is calculated based on the validity of the initiator's contribution proofs during the historical anchoring period, as well as the stability of node parameters reflected by communication time costs.

Images