A federated learning dynamic revocation and parameter-level forgetting method and system based on a healthy digital identity

Abstract

The application provides a health digital identity-based federated learning dynamic revocation and parameter-level forgetting method and system, and belongs to the technical field of computer federated learning. Trend feature sequences are extracted from health monitoring event streams; whether the dynamic revocation trigger condition is met is judged according to the continuous change slope and fluctuation amplitude in the trend feature sequences, if the condition is met, a revocation start instruction is generated to locate all participating nodes associated with the health digital identity, and a to-be-revoked parameter set is parsed out, the change severity value is calculated for the acceleration and deceleration change sections respectively, the parameters in the to-be-revoked parameter set are grouped and divided according to the type, and the corresponding forgetting strategy is configured, the corresponding parameter-level forgetting operation is performed on each parameter group, and the revocation timestamp and the type label of the forgotten parameters are written in the digital identity revocation record table. The application can adaptively configure differentiated parameter forgetting strategies according to health trend changes, and realize accurate and efficient dynamic revocation and parameter-level forgetting.

Description

Technical Field

[0001] This invention provides a method and system for dynamic revocation and parameter-level forgetting in federated learning based on health digital identity, belonging to the field of computer federated learning technology. Background Technology

[0002] With the widespread adoption of wearable devices and remote health monitoring technologies, digital health identities have become a crucial link between user physiological data and federated learning systems. Federated learning, as a distributed machine learning paradigm, enables model updates through collaboration among multiple participating nodes without requiring centralized collection of raw data. In health monitoring scenarios, each user's digital health identity generates a continuously changing sequence of health indicators over time, reflecting the dynamic evolution of various physiological parameters such as heart rate, blood pressure, and respiratory rate. Federated learning systems typically maintain a set of associated parameters for each digital health identity, distributed across different participating nodes, to characterize the identity's health features. However, when a user stops using the service, cancels their account, or data protection policies require the deletion of specific identity information, the system needs to perform a user revocation operation and thoroughly clean up the parameters associated with that identity.

[0003] Existing user withdrawal methods in federated learning mostly employ global model retraining or periodic node exit mechanisms, which have significant drawbacks.

[0004] On the one hand, global retraining consumes significant computational resources and communication bandwidth, and cannot perform fine-grained, real-time revocation of individual health digital identities. On the other hand, traditional parameter deletion methods typically employ a uniform forgetting strategy for all parameters, ignoring the varying sensitivities of different parameters to changes in health trends. When a user's health indicators show an accelerating deterioration or rapid improvement trend, existing methods cannot dynamically adjust the timing of revocation and the intensity of forgetting based on trend changes. This results in the residual physiological characteristics corresponding to revoked user identities potentially being indirectly restored through model inference. Furthermore, the lack of hierarchical processing of parameter sensitivity during parameter-level forgetting means that some highly sensitive parameters are not eliminated in a timely and thorough manner, while low-sensitivity parameters may be overprocessed, affecting the service quality of other users in the federated learning system. Summary of the Invention

[0005] To address the aforementioned issues, a dynamic revocation and parameter-level forgetting method based on health digital identities in federated learning needs to be designed. This method should be able to determine revocation trigger conditions in real time based on the trend feature sequence of health digital identities and configure differentiated forgetting strategies for different types of parameters. The method should be able to extract key features reflecting health change trends from the health monitoring event stream, and accordingly classify related parameters into different sensitivity levels: high-sensitivity parameters should be configured with instantaneous complete zeroing, medium-sensitivity parameters with gradual decay, and low-sensitivity parameters with linear descent. Simultaneously, the method should also possess adaptive capabilities, dynamically adjusting parameter grouping and strategy configuration according to the acceleration or deceleration of health trends to avoid delayed or over-responding to extreme health changes. Through this design, accurate revocation and parameter-level forgetting of health digital identities can be achieved while ensuring privacy compliance, and the adverse impact on the overall performance of the federated learning system and the quality of remaining user models can be reduced.

[0006] This application provides a method for dynamic revocation and parameter-level forgetting of federated learning based on health digital identity, including the following steps: S1: Collect the health monitoring event stream corresponding to the health digital identity, and extract the trend feature sequence of the health digital identity from the health monitoring event stream; S2: Based on the continuous change slope and fluctuation amplitude in the trend feature sequence, determine whether the triggering condition for dynamic cancellation is met within the current time window. If it is met, generate a cancellation start command. S3: In response to the revocation initiation command, locate all participating nodes associated with the health digital identity in the federated learning architecture, and parse out the set of parameters to be revoked; S4: Calculate the degree of change for the accelerated and decelerated change segments in the trend feature sequence, group each parameter type in the set of parameters to be undone, and configure the corresponding forgetting strategy. S5: Perform parameter-level forgetting operation on each parameter in the sensitive parameter group according to the forgetting strategy; S6: After completing the parameter-level forgetting, write the timestamp of this revocation and the type marker of the forgotten parameters into the digital identity cancellation record table corresponding to the health digital identity.

[0007] Furthermore, the triggering condition for dynamic cancellation in S2 is determined in the following way: Extract health indicator values ​​for three consecutive time points from the trend feature sequence, calculate the first slope of change from the first time point to the second time point, and the second slope of change from the second time point to the third time point. If both the first slope of change and the second slope of change are greater than a preset positive threshold and the difference between the first slope of change and the second slope of change is less than a preset stable deviation, then it is determined to be a continuous positive acceleration trend and dynamic cancellation is triggered. If both the first change slope and the second change slope are less than a preset negative threshold and the difference between the first change slope and the second change slope is less than a preset stable deviation, then it is determined to be a continuous negative acceleration trend and dynamic cancellation is triggered. If the product of the first slope and the second slope is negative and the fluctuation amplitude of the trend feature sequence exceeds the preset violent fluctuation threshold in the three consecutive time points, it is determined to be a trend reversal fluctuation and dynamic cancellation is triggered.

[0008] Furthermore, S3 includes: Read the identification code of the health digital identity, input it into the pre-established node mapping directory, and obtain the list of first participating nodes through hash lookup; Then, based on the node association frequency extracted from the historical interaction records of the health digital identity, cold nodes with association frequencies lower than a preset frequency threshold are filtered out from the first participating node list to obtain the second participating node list; For each participating node in the second participating node list, a parameter query request is sent to that participating node, the parameter index table returned by the participating node is received, and the parameter storage address marked with the health digital identity tag is selected from the parameter index table to form the set of parameters to be revoked.

[0009] Furthermore, S4 performs the following operations: The trend feature sequence is segmented and identified. A sliding time window is set, and the rate of change between each pair of adjacent time points within the sliding time window is calculated to obtain multiple rate of change values. If the rate of change values ​​increases sequentially and the difference between adjacent rate of change values ​​is greater than the first increment threshold, then the sliding time window is marked as an accelerated change segment. If these multiple rate-of-change values ​​decrease sequentially and the difference between adjacent rate-of-change values ​​is greater than the first decreasing threshold, then the sliding time window is marked as a deceleration segment. For each identified acceleration segment, the difference between the maximum and minimum rate of change within that segment is taken as the first severity value of the acceleration segment. For each identified deceleration segment, the difference between the absolute value of the maximum and the absolute value of the minimum rate of change within that segment is taken as the second severity value of the deceleration segment. Iterate through each parameter type group in the set of parameters to be revoked, which contains multiple parameter instances and is associated with parameter sensitivity labels. Add up the parameter sensitivity label values ​​of all parameter instances in the parameter type group and divide by the total number of instances in the parameter type group to obtain the average sensitivity score of the parameter type group. Multiply the average sensitivity score by the first severity value to obtain the sensitivity weight under accelerated change, and multiply the average sensitivity score by the second severity value to obtain the sensitivity weight under decelerated change. Take the larger value between the sensitivity weight under accelerated change and the sensitivity weight under decelerated change as the comprehensive sensitivity weight for this parameter type group.

[0010] Further, S5 includes: Iterate through all parameter storage addresses in the high-sensitivity parameter group. For each storage address, read the current parameter value and determine whether the current parameter value is equal to the preset initial parameter value. If not, set all binary bits in the storage address to logic 0 and update the checksum of the storage address to the checksum corresponding to all 0s. For each parameter in the sensitive parameter group, obtain the current value of the parameter, read the attenuation count value associated with the parameter. The attenuation count value is initially zero. If the attenuation count value is less than the maximum attenuation count, divide the current value of the parameter by the attenuation base to obtain an intermediate value. Round the intermediate value according to the data type precision of the parameter to obtain an updated value. Write the updated value back to the storage address of the parameter and increment the attenuation count value by one. If the attenuation count value has reached the maximum attenuation count, skip further attenuation operations for the parameter and mark it as completely forgotten in the parameter's metadata.

[0011] Furthermore, in step S4, the degree of change is calculated for the accelerated and decelerated change segments in the trend feature sequence. Based on the degree of change, each parameter type in the set of parameters to be withdrawn is grouped into a high-sensitivity parameter group, a medium-sensitivity parameter group, and a low-sensitivity parameter group. An instantaneous zeroing forgetting strategy is configured for the high-sensitivity parameter group, an exponential decay forgetting strategy is configured for the medium-sensitivity parameter group, and a linear bottoming forgetting strategy is configured for the low-sensitivity parameter group.

[0012] Furthermore, before configuring the forgetting strategy, dynamic reclassification operations are performed for the following parameter type groups: Obtain the historical revocation record table of the health digital identity, which records the trend feature sequence fragments of the five most recent revocation events and the forgetting strategy identifier configured for each parameter type group at that time; The first intensity value of the acceleration change segment and the second intensity value of the deceleration change segment calculated in the current S4 are compared with the historical intensity values ​​of the corresponding trend feature sequence segments in the historical cancellation record table. If the current first intensity value exceeds twice any historical intensity value, the current acceleration change segment is determined to be an abnormal acceleration. If the current second intensity value exceeds twice any historical intensity value, the current deceleration change segment is determined to be an abnormal deceleration. When it is determined to be an abnormal acceleration or deceleration, the duration of the acceleration or deceleration segment in the current trend feature sequence is extracted. The duration is measured by the number of time points. If the duration is greater than a preset length threshold, a reclassification request is initiated. The reclassification request carries the identifier of the current trend feature sequence. In response to the reclassification request, all parameter type groups in the set of parameters to be revoked are sorted from high to low according to their average sensitivity scores. The top 20% of the parameter type groups are forcibly assigned to the high-sensitivity parameter group and configured with an instantaneous zeroing forgetting strategy. The bottom 20% of the parameter type groups are forcibly assigned to the low-sensitivity parameter group and configured with a linear bottoming forgetting strategy. The remaining parameter type groups are divided into high-sensitivity, medium-sensitivity, and low-sensitivity parameter groups and configured with corresponding strategies.

[0013] Furthermore, the health monitoring event stream corresponding to the health digital identity in S1 is serialized event data received in real time from the wearable device interface, and the extraction process of the trend feature sequence is specifically as follows: For each event in the health monitoring event stream, extract the event occurrence timestamp and raw health indicator readings, and arrange the raw health indicator readings into a raw reading sequence. In the original reading sequence, a sampling point is taken every fixed number of events to obtain a sampling point sequence. The difference between adjacent sampling points in the sampling point sequence is calculated, and the difference is divided by the time interval between adjacent sampling points to obtain the rate of change sequence. The rate of change sequence is filtered by moving average, with the window size of the moving average set to three rate of change values ​​to obtain a smoothed rate of change sequence. Then, three consecutive smoothed rate of change values ​​in the smoothed rate of change sequence are combined into a trend feature vector. All trend feature vectors are arranged in chronological order to form a trend feature sequence.

[0014] According to a second aspect of the present invention, the present invention claims protection for a federated learning dynamic revocation and parameter-level forgetting system based on health digital identity, comprising: One or more processors; A memory storing one or more programs that, when executed by one or more processors, enable the processors to implement the described method for dynamic revocation and parameter-level forgetting in federated learning based on health digital identity.

[0015] This invention provides a method and system for dynamic revocation and parameter-level forgetting based on health digital identity in federated learning, belonging to the field of computer federated learning technology. It extracts trend feature sequences from a health monitoring event stream; determines whether dynamic revocation triggering conditions are met based on the continuous change slope and fluctuation amplitude in the trend feature sequences; if met, it generates a revocation initiation command, locates all participating nodes associated with the health digital identity, and parses the set of parameters to be revoked. It calculates the degree of change for acceleration and deceleration segments, groups the parameter types in the set of parameters to be revoked according to this, configures corresponding forgetting strategies, performs corresponding parameter-level forgetting operations on each parameter group, and writes the revocation timestamp and the type marker of the forgotten parameters into the digital identity cancellation record table. This invention can adaptively configure differentiated parameter forgetting strategies according to changes in health trends, achieving accurate and efficient dynamic revocation and parameter-level forgetting. Attached Figure Description

[0016] Figure 1 This application, as claimed in the embodiments of the present invention, provides a flowchart of a federated learning dynamic revocation and parameter-level forgetting method based on health digital identity; Figure 2 This application, as claimed in the embodiments of the present invention, provides a second flowchart of a federated learning dynamic revocation and parameter-level forgetting method based on health digital identity; Figure 3 This application, as claimed in the embodiments of the present invention, provides a third workflow diagram of a federated learning dynamic revocation and parameter-level forgetting method based on health digital identity. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] The technical solutions disclosed in the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] According to a first embodiment of the present invention, the present invention claims protection for a method for dynamic revocation and parameter-level forgetting in federated learning based on health digital identity, referring to... Figure 1 This includes the following steps: S1: Collect the health monitoring event stream corresponding to the health digital identity, and extract the trend feature sequence of the health digital identity from the health monitoring event stream; S2: Based on the continuous change slope and fluctuation amplitude in the trend feature sequence, determine whether the triggering condition for dynamic cancellation is met within the current time window. If it is met, generate a cancellation start command. S3: In response to the revocation initiation command, locate all participating nodes associated with the health digital identity in the federated learning architecture, and parse out the set of parameters to be revoked; S4: Calculate the degree of change for the accelerated and decelerated change segments in the trend feature sequence, group each parameter type in the set of parameters to be undone, and configure the corresponding forgetting strategy. S5: Perform parameter-level forgetting operation on each parameter in the sensitive parameter group according to the forgetting strategy; S6: After completing the parameter-level forgetting, write the timestamp of this revocation and the type marker of the forgotten parameters into the digital identity cancellation record table corresponding to the health digital identity.

[0020] In this embodiment, the health digital identity corresponds to a user who wears an ECG monitoring bracelet for an extended period. The bracelet generates three health indicators—ECG characteristic value, heart rate variability, and respiratory rate—every five seconds, and packages these data into a health monitoring event. All events form a continuous health monitoring event stream in chronological order. First, step S1 is executed: the health monitoring event stream is sequentially read from the wearable device's receive buffer, and the timestamp and values ​​of the three health indicators are extracted for each event. An internal circular buffer with a capacity of one thousand events is maintained. New events are appended whenever the buffer is not full; when the buffer is full, the latest event overwrites the oldest event. Then, the most recent two hundred events are retrieved from the buffer in chronological order. For the heart rate variability indicator in each event, the difference between this indicator and the heart rate variability of the previous event is calculated to obtain a difference sequence. This difference sequence is then summed using a seven-point shift algorithm to obtain a trend feature sequence. Each element in this trend feature sequence represents the comprehensive trend direction of changes in the health digital identity within a time window.

[0021] The trend feature sequence is fed into the trend analysis module, which maintains a sliding window with a width of ten trend feature values. Within the window, the difference between all adjacent trend feature values ​​is calculated to obtain a series of continuously changing slopes. Simultaneously, the difference between the maximum and minimum trend feature values ​​within the window is calculated as the fluctuation amplitude. Three dynamic cancellation trigger conditions are preset: the first condition is five consecutive positive slopes, each greater than 0.3; the second condition is five consecutive negative slopes, each less than -0.3; and the third condition is that the fluctuation amplitude of three consecutive windows is greater than 1.2. The calculated consecutive slopes for the current window are 0.41, 0.43, 0.44, 0.42, and 0.45, satisfying the first condition, thus generating a cancellation start command. This command contains a unique identifier for the health digital identity and a system timestamp of the trigger time.

[0022] In response to the revocation initiation command, the system accesses the global node registry center, which stores a mapping table. The key is the health digital identity identifier, and the value is the network address of a group of participating nodes. A hash lookup is used to obtain the first list of participating nodes, consisting of twelve nodes. Then, the system queries the historical interaction record database for this health digital identity, which records the frequency of data exchanges between each node and this identity over the past thirty days. A frequency threshold of five times is set, and nodes with fewer than five interactions are removed from the first list, resulting in the second list of participating nodes, leaving eight nodes. Parameter query requests are sent to these eight nodes in parallel. Each node returns a parameter index table, where each row contains the parameter name, parameter storage address, and an attribution flag field. If the attribution flag field equals the health digital identity identifier, the parameter address is added to the set of parameters to be revoked. The final set of parameters to be revoked contains three parameter type groups: the first group is ECG feature convolution kernel parameters, with 256 parameter instances; the second group is heart rate variability fully connected layer weight parameters, with 128 parameter instances; and the third group is respiratory rate bias parameters, with over 60 parameter instances.

[0023] The severity of change is calculated based on the acceleration and deceleration segments in the trend feature sequence. In this embodiment, the most recent ten values ​​of the trend feature sequence exhibit an acceleration followed by deceleration pattern: the differences between the first five values ​​are 0.1, 0.15, 0.2, and 0.25, representing the acceleration segment; the differences between the last five values ​​are 0.25, 0.2, 0.15, and 0.1, representing the deceleration segment. The severity of the acceleration segment is calculated as the difference between the maximum rate of change (0.25) and the minimum rate of change (0.1), which is 0.15; the severity of the deceleration segment is calculated as the difference between the absolute value of the maximum rate of change (0.25) and the absolute value of the minimum rate of change (0.1), which is 0.15. Sensitivity labels are pre-assigned to each parameter type group: the sensitivity label for the ECG feature convolution kernel group is 95, for the heart rate variability fully connected layer group is 55, and for the respiratory rate bias group is 20. The average sensitivity scores for each group are calculated to be 95, 55, and 20, respectively. The average sensitivity score is multiplied by the severity value to obtain the overall sensitivity weight: 14.25 for the first group, 8.25 for the second group, and 3.00 for the third group. A first threshold range of 0 to 4 and a second threshold range of 10 to 20 are pre-set. Therefore, the first group (14.25) falls into the second range and is classified as a high-sensitivity parameter group; the second group (8.25) is between 4 and 10 and is classified as a medium-sensitivity parameter group; the third group (3.00) falls into the first range and is classified as a low-sensitivity parameter group. The following strategies are then configured: a rapid zeroing forgetting strategy is used for the high-sensitivity parameter group, where each parameter is directly written to zero; an exponential decay forgetting strategy is used for the medium-sensitivity parameter group, where each parameter decays to half its current value each time, with each decay occurring after a health indicator update cycle; and a linear bottoming forgetting strategy is used for the low-sensitivity parameter group, where each parameter is subtracted by a fixed step size of 0.01 each time until a preset lower limit of -1.0 is reached.

[0024] According to the configuration, the forgetting operation is performed. For the 256 ECG feature convolution kernel parameters in the high-sensitivity parameter group, their storage addresses are accessed one by one, each binary bit in the storage unit is set to logic 0, and the checksum of the address is recalculated to the value corresponding to all 0. For the 128 heart rate variability fully connected layer weight parameters in the medium-sensitivity parameter group, the current value of each parameter is read, divided by 2 to obtain a new value, rounded according to the floating-point single-precision format, and written back. At the same time, the decay count is recorded and incremented by one in the metadata area of ​​the parameter. For the more than 60 respiratory rate bias parameters in the low-sensitivity parameter group, the current value is read and 0.01 is subtracted. It is determined whether it is greater than or equal to -1.0. If it is, the new value is written back. If it has reached -1.0, it remains unchanged.

[0025] After all forgetting operations are completed, the digital identity deregistration record table in persistent storage is accessed. This table uses the health digital identity identifier as the primary key and contains two columns: a list of revocation timestamps and a list of forgotten parameter type tags. In this operation, the current system time, April 25, 2026, 10:30:15 AM, is appended to the revocation timestamp list, and the parameter type tags—ECG feature convolution kernel, heart rate variability fully connected layer weights, and respiratory rate bias—are written to the forgotten parameter type tag column. Simultaneously, a status field with the value "Completed" is written to this record table. At this point, the entire dynamic revocation and parameter-level forgetting process is complete.

[0026] Furthermore, the triggering condition for dynamic cancellation in S2 is determined in the following way: Extract health indicator values ​​for three consecutive time points from the trend feature sequence, calculate the first slope of change from the first time point to the second time point, and the second slope of change from the second time point to the third time point. If both the first slope of change and the second slope of change are greater than a preset positive threshold and the difference between the first slope of change and the second slope of change is less than a preset stable deviation, then it is determined to be a continuous positive acceleration trend and dynamic cancellation is triggered. If both the first change slope and the second change slope are less than a preset negative threshold and the difference between the first change slope and the second change slope is less than a preset stable deviation, then it is determined to be a continuous negative acceleration trend and dynamic cancellation is triggered. If the product of the first slope and the second slope is negative and the fluctuation amplitude of the trend feature sequence exceeds the preset violent fluctuation threshold in the three consecutive time points, it is determined to be a trend reversal fluctuation and dynamic cancellation is triggered.

[0027] In this embodiment, health indicator values ​​at three consecutive time points are extracted from the trend feature sequence. These three time points are denoted as T1, T2, and T3, with corresponding heart rate variability values ​​of 72.3, 74.1, and 76.2, respectively. The first slope of change is calculated: the change from T1 to T2 is 1.8, with a time interval of five seconds, and the slope is 0.36; the second slope of change is calculated: the change from T2 to T3 is 2.1, with a time interval of five seconds, and the slope is 0.42. A positive threshold of 0.30 and a stability deviation of 0.10 are preset. Since both slopes are greater than 0.30, and the difference between them is 0.06, which is less than the stability deviation of 0.10, a continuous positive acceleration trend is identified, triggering dynamic cancellation.

[0028] Another scenario: If the heart rate variability values ​​for T1, T2, and T3 are 80.5, 77.2, and 73.8 respectively, then the first slope of change is -0.66, the second slope of change is -0.68, the preset negative threshold is -0.30, and the stability deviation is 0.10. If both slopes are less than -0.30, and the absolute value of the difference is 0.02 (less than 0.10), it is determined to be a continuous negative acceleration trend, triggering dynamic cancellation.

[0029] The third scenario: Taking values ​​of 75.0, 78.5, and 74.0, the first slope of the change is 0.70 (positive), and the second slope is -0.90 (negative). The product is -0.63, which is negative. The fluctuation amplitude is calculated as follows: the maximum value is 78.5, the minimum value is 74.0, and the difference is 4.5. The preset threshold for severe fluctuation is 3.0. Since 4.5 is greater than 3.0, it is determined to be a trend reversal fluctuation, triggering dynamic cancellation.

[0030] In the implementation, a circular queue of length one hundred is maintained, with each element being (timestamp, health metric value). Upon receiving a new event, the last element is popped from the queue, and the new event is pushed to the front. Then, starting from the front, three consecutive elements are taken and the above conditions are checked. If none of the three conditions are met, no cancellation start command is generated, and the system continues to wait for the next event. If any condition is met, a cancellation start command is immediately generated, and the trigger condition type (positive acceleration, negative acceleration, or reverse fluctuation) is encoded into the command so that different parameter-sensitive groups in subsequent steps can reference this type of information.

[0031] Furthermore, S3 includes: Read the identification code of the health digital identity, input it into the pre-established node mapping directory, and obtain the list of first participating nodes through hash lookup; Then, based on the node association frequency extracted from the historical interaction records of the health digital identity, cold nodes with association frequencies lower than a preset frequency threshold are filtered out from the first participating node list to obtain the second participating node list; For each participating node in the second participating node list, a parameter query request is sent to that participating node, the parameter index table returned by the participating node is received, and the parameter storage address marked with the health digital identity tag is selected from the parameter index table to form the set of parameters to be revoked.

[0032] In this embodiment, the identity code of the health digital identity is read. This identity code is a 128-bit UUID string, such as 3F2504E0-4F89-11D3-9A0C-0305E82C3301. This string is input into the node mapping directory. The node mapping directory is implemented by a distributed consistent hash ring, with each participating node responsible for a hash space. The hash value of the UUID is calculated, and the first server is searched clockwise on the hash ring to obtain the first participating node. Then, the search continues clockwise until all nodes related to the identity are collected. In actual operation, each entry in the mapping directory stores the identity code and a list of node network addresses. The list of the first participating nodes is obtained directly through exact matching, totaling twelve nodes.

[0033] Next, the historical interaction record table is queried. This table is stored in the central coordination server and its structure is: identity code, node ID, interaction timestamp, and interaction data volume. The number of interactions between this identity and each node in the past thirty days is counted, and the following data is obtained: Node A: 23 times, Node B: 18 times, Node C: 7 times, Node D: 4 times, Node E: 2 times, Node F: 1 time, and the remaining nodes have 0 interactions; the preset frequency threshold is 5 times. Therefore, nodes A, B, and C are retained, while nodes D, E, and F, as well as nodes with 0 interactions, are filtered out, resulting in the second list of participating nodes. The remaining three nodes, A, B, and C, are actually only three active nodes after filtering out the original twelve nodes in this example.

[0034] Then, parameter query requests are sent to nodes A, B, and C respectively. The message format of this request includes: a request type field with a value of 0x03 indicating a parameter query, an identity code field, and a request sequence number field. Upon receiving the request, each node searches its local parameter storage area for all parameter metadata. Each parameter metadata contains an attribution identity list field, which is a variable-length array storing all healthy digital identity identifiers that have contributed to that parameter. The node iterates through all parameter metadata; if the attribution identity list contains the target identity code, it combines the parameter's name, storage address, data type, current numerical version number, etc., into a record and adds it to the node's parameter index table. The node then encapsulates the index table in a response message and returns it.

[0035] After receiving responses from the three nodes, all records are merged and deduplicated. If the same parameter is stored on multiple nodes, it is distinguished by the node identifier. Finally, a set of parameters to be revoked is formed. This set contains three parameter type groups, and each group contains multiple parameter instances. The complete location information of each parameter instance is the node address, storage address offset, and parameter length.

[0036] Furthermore, referring to Figure 2 S4 performs the following operations: The trend feature sequence is segmented and identified. A sliding time window is set, and the rate of change between each pair of adjacent time points within the sliding time window is calculated to obtain multiple rate of change values. If the rate of change values ​​increases sequentially and the difference between adjacent rate of change values ​​is greater than the first increment threshold, then the sliding time window is marked as an accelerated change segment. If these multiple rate-of-change values ​​decrease sequentially and the difference between adjacent rate-of-change values ​​is greater than the first decreasing threshold, then the sliding time window is marked as a deceleration segment. For each identified acceleration segment, the difference between the maximum and minimum rate of change within that segment is taken as the first severity value of the acceleration segment. For each identified deceleration segment, the difference between the absolute value of the maximum and the absolute value of the minimum rate of change within that segment is taken as the second severity value of the deceleration segment. Iterate through each parameter type group in the set of parameters to be revoked, which contains multiple parameter instances and is associated with parameter sensitivity labels. Add up the parameter sensitivity label values ​​of all parameter instances in the parameter type group and divide by the total number of instances in the parameter type group to obtain the average sensitivity score of the parameter type group. Multiply the average sensitivity score by the first severity value to obtain the sensitivity weight under accelerated change, and multiply the average sensitivity score by the second severity value to obtain the sensitivity weight under decelerated change. Take the larger value between the sensitivity weight under accelerated change and the sensitivity weight under decelerated change as the comprehensive sensitivity weight for this parameter type group.

[0037] In this embodiment, segmented identification is performed on the trend feature sequence, maintaining a trend feature value array of fixed length ten. Each trend feature value is a floating-point number representing the average rate of change of the health indicator within a five-second window. The length of the sliding time window is set to five consecutive time points, corresponding to multiple adjacent rates of change. Starting from the first time point, the trend feature values ​​corresponding to time points P1, P2, P3, P4, and P5 are extracted as 0.12, 0.15, 0.19, 0.24, and 0.30, respectively. Multiple rates of change are calculated: 0.03, 0.04, 0.05, and 0.06. These values ​​increase sequentially, and the adjacent differences are 0.01, 0.01, and 0.01, all greater than the first increment threshold preset to 0.005; therefore, this sliding time window is marked as an accelerated change segment. Take another set of values ​​from P6 to P10: 0.32, 0.30, 0.27, 0.23, 0.18, with change rates of -0.02, -0.03, -0.04, and -0.05 respectively. They decrease and the absolute value of the difference is greater than the first decreasing threshold preset to 0.005, and are marked as deceleration change segments.

[0038] For the identified acceleration segment, the difference of 0.03 between the maximum rate of change (0.06) and the minimum rate of change (0.03) within that segment is taken as the first severity value. For the deceleration segment, the difference of 0.03 between the absolute value of the maximum rate of change (0.05) and the absolute value of the minimum rate of change (0.02) is taken as the second severity value.

[0039] Then, iterate through each parameter type group in the set of parameters to be revoked. In this embodiment, there are five parameter type groups: convolutional layer weight group, batch normalized scaling group, batch normalized offset group, fully connected layer weight group, and output layer bias group. Each parameter instance in each group is pre-assigned a parameter sensitivity label during the production stage, with label values ​​ranging from 0 to 100. The specific assignment rules are as follows: labels for the convolutional layer weight group are randomly assigned between 90 and 100; labels for the batch normalized scaling group are between 70 and 85; labels for the batch normalized offset group are between 60 and 75; labels for the fully connected layer weight group are between 40 and 60; and labels for the output layer bias group are between 10 and 30. The sum of the sensitivity label values ​​for all parameter instances in each group is calculated and then divided by the total number of instances to obtain the average sensitivity score. For example, the convolutional layer weight group has 128 instances, with a total label sum of 12160 and an average score of 95; the output layer bias group has 8 instances, with a total label sum of 160 and an average score of 20.

[0040] Multiplying the average sensitivity score by the first intensity value of 0.03 yields the sensitivity weight under accelerated change, and multiplying it by the second intensity value of 0.03 yields the sensitivity weight under decelerated change. The larger of the two values ​​is taken as the comprehensive sensitivity weight for this parameter type group. For convolutional layer weight grouping: the accelerated sensitivity weight is 95. 0.03 = 2.85, the deceleration sensitivity weights are the same, the sum is 2.85, for output layer bias grouping: the sum is 20. 0.03 = 0.6. The first threshold interval is pre-set to [0, 1.0], and the second threshold interval is [2.0, 10.0]. Therefore, a combined sensitivity weight of 0.6 falls into the first interval and is classified as a low-sensitivity parameter group; a combined sensitivity weight of 2.85 falls into the second interval and is classified as a high-sensitivity parameter group; if any group's combined sensitivity weight falls between 1.0 and 2.0, it is classified as a medium-sensitivity parameter group.

[0041] When configuring the strategy, for each parameter type group of the highly sensitive parameter group, such as the convolutional layer weight group, a strategy configuration structure is created. This structure includes: setting the zeroing operation flag to logical true and setting the number of delayed execution cycles to zero, meaning that zeroing is performed immediately at the start of S5. In addition, a post-zeroing verification flag is set, requiring that the cyclic redundancy check value of the storage block be recalculated and written to the verification area after zeroing.

[0042] For the moderately sensitive parameter group, configure an exponential decay forgetting strategy; calculate the decay period length: query the half-life value of the health indicator for this health digital identity. This value is determined by the rate of decline of the indicator in the health monitoring history. For example, if the ECG characteristic decreases by half within fifteen minutes after stopping exercise, the half-life value is 900 seconds. Set the decay period length to 900 seconds. Set the maximum number of decays to ten. Set the decay base to half, meaning that after each decay, the value becomes 0.5 times the original value.

[0043] For low-sensitivity parameter groups, a linear bottoming-out forgetting strategy is configured. The minimum parameter adjustment unit for this parameter type group is queried. For example, the output layer bias parameter's numerical type is a 32-bit floating-point number, and the minimum representable step size is 1.1920929e-7. However, for practical operability, the linear descent step size is set to 0.0001. Then, the parameter value range length is calculated: the allowed value range for this group's parameters is -1.0 to 1.0, and the range length is 2.0. The total number of descent steps is 2.0 divided by 0.0001, which equals 20,000 steps, rounded down to 20,000. The descent interval is set to the health indicator update interval for this health digital identity, i.e., five seconds. This means that a descent operation is performed every five seconds, decreasing by 0.0001 each time, until 20,000 operations are completed or the parameter value reaches the lower limit of -1.0.

[0044] Finally, the mapping relationship between group identifiers and policy configurations is stored in a forgetting policy table for S5 to call.

[0045] Further, S5 includes: Iterate through all parameter storage addresses in the high-sensitivity parameter group. For each storage address, read the current parameter value and determine whether the current parameter value is equal to the preset initial parameter value. If not, set all binary bits in the storage address to logic 0 and update the checksum of the storage address to the checksum corresponding to all 0s. For each parameter in the sensitive parameter group, obtain the current value of the parameter, read the attenuation count value associated with the parameter. The attenuation count value is initially zero. If the attenuation count value is less than the maximum attenuation count, divide the current value of the parameter by the attenuation base to obtain an intermediate value. Round the intermediate value according to the data type precision of the parameter to obtain an updated value. Write the updated value back to the storage address of the parameter and increment the attenuation count value by one. If the attenuation count value has reached the maximum attenuation count, skip further attenuation operations for the parameter and mark it as completely forgotten in the parameter's metadata.

[0046] In this embodiment, each parameter in the highly sensitive parameter group is directly zeroed out. The highly sensitive parameter group contains convolutional layer weight groups, which are stored in the memory address range 0x1000 to 0x2000 of node A. The storage address of each parameter within this range is traversed. For the weight value stored at address 0x1000, the current value is 0.873. The system checks if the value is equal to the initial parameter value. The initial parameter value is set to a small random number sampled from a normal distribution during model initialization and is not equal to zero. Since 0.873 is not equal to the preset initial parameter value, such as 0.031, a zeroing operation is performed: the underlying memory write interface is called to write all four bytes of single-precision floating-point numbers starting at address 0x1000 into binary 0x00000000; at the same time, the checksum of this storage address is updated. This checksum is stored in an independent area starting at address 0x2000, with each parameter corresponding to one byte of checksum; after the zeroing operation is completed, the value at the target address is read out and compared with all zeros. If they match, a new checksum is calculated as the XOR value of all zeros, 0x00, and written into the checksum area. In addition, a reset timestamp and a reset reason code are written to the metadata area of ​​the storage address. The metadata area is an independent array of structures, each element of which contains the parameter address, the reset timestamp, and the reset reason code. The reset timestamp is taken as the high-precision counter value of the current time, and the reset reason code is set to 0x01, which corresponds to the strategy number of the instantaneous zeroing forgetting strategy.

[0047] For all 256 parameters in the high-sensitivity parameter group, perform the above operations sequentially. If the current value of a parameter is already zero, skip the zeroing operation but still write a timestamp and reason code to record this processing.

[0048] Next, for each parameter in the medium-sensitive parameter group, a halving operation is performed on the value of each parameter according to the preset decay period. The medium-sensitive parameter group is a weight group of the fully connected layer for heart rate variability, stored in the memory of node B. Each parameter is associated with a decay count value, which is stored in the extended field of the parameter metadata and is initially zero. The current value of 0.512 at address 0x3000 of the first parameter is retrieved, and the decay count value is read as 0. The maximum number of decays is preset to ten, and the decay count value of 0 is less than 10. The value is divided by two: 0.512 ÷ 2 = 0.256. Since the data uses single-precision floating-point numbers, banker's rounding is used for rounding. 0.256 is precisely represented as 0.256000012 in single precision. This value is directly written back, and then the decay count value is updated to 1. The time of this decay execution is recorded with a time precision of milliseconds. For the second parameter, the current value is -0.384, and after decay it is -0.192. The same write-back and count update are performed. Each parameter maintains an independent decay count value, and the decay progress of different parameters does not affect each other. When the decay count value of a parameter reaches ten times, further decay operations for that parameter are skipped, and a flag of complete forgetting with a value of 1 is written into the parameter's metadata.

[0049] Specifically, the decay period is controlled by the scheduler. When S5 executes, it does not wait for a period before executing the next decay, but immediately executes the first decay. The preset decay period length represents the time interval between two decays, but S5 only performs the decay operation once. Subsequent decays are completed by a background periodic task. This task scans all parameters in the sensitive parameter group every decay period, and continues to halve the decay count for parameters whose decay count has not reached the maximum number of times.

[0050] Furthermore, in step S4, the degree of change is calculated for the accelerated and decelerated change segments in the trend feature sequence. Based on the degree of change, each parameter type in the set of parameters to be withdrawn is grouped into a high-sensitivity parameter group, a medium-sensitivity parameter group, and a low-sensitivity parameter group. An instantaneous zeroing forgetting strategy is configured for the high-sensitivity parameter group, an exponential decay forgetting strategy is configured for the medium-sensitivity parameter group, and a linear bottoming forgetting strategy is configured for the low-sensitivity parameter group.

[0051] Furthermore, referring to Figure 3 Before configuring the forgetting strategy, dynamic reclassification operations are performed for the following parameter types: Obtain the historical revocation record table of the health digital identity, which records the trend feature sequence fragments of the five most recent revocation events and the forgetting strategy identifier configured for each parameter type group at that time; The first intensity value of the acceleration change segment and the second intensity value of the deceleration change segment calculated in the current S4 are compared with the historical intensity values ​​of the corresponding trend feature sequence segments in the historical cancellation record table. If the current first intensity value exceeds twice any historical intensity value, the current acceleration change segment is determined to be an abnormal acceleration. If the current second intensity value exceeds twice any historical intensity value, the current deceleration change segment is determined to be an abnormal deceleration. When it is determined to be an abnormal acceleration or deceleration, the duration of the acceleration or deceleration segment in the current trend feature sequence is extracted. The duration is measured by the number of time points. If the duration is greater than a preset length threshold, a reclassification request is initiated. The reclassification request carries the identifier of the current trend feature sequence. In response to the reclassification request, all parameter type groups in the set of parameters to be revoked are sorted from high to low according to their average sensitivity scores. The top 20% of the parameter type groups are forcibly assigned to the high-sensitivity parameter group and configured with an instantaneous zeroing forgetting strategy. The bottom 20% of the parameter type groups are forcibly assigned to the low-sensitivity parameter group and configured with a linear bottoming forgetting strategy. The remaining parameter type groups are divided into high-sensitivity, medium-sensitivity, and low-sensitivity parameter groups and configured with corresponding strategies.

[0052] In this embodiment, a historical revocation record table of the health digital identity is obtained. This table is stored in a central database and records information for the five most recent revocation events. Each record includes: the revocation event number, the first intensity value of the accelerated change segment calculated at that time, the second intensity value of the deceleration change segment, and the forgetting strategy identifier configured for each parameter type group at that time, such as strategy identifier 0 indicating instantaneous zeroing, 1 indicating exponential decay, and 2 indicating linear bottoming. For example, in the five most recent events: the first event: first intensity value 0.02, second intensity value 0.018, strategy identifiers: convolutional layer: 0, batch normalization: 1, output layer: 2; the second event: first intensity value 0.025, second intensity value 0.022, similar strategy identifiers; the third event: first intensity value 0.03, second intensity value 0.028; the fourth event: first intensity value 0.027, second intensity value 0.025; the fifth event: first intensity value 0.031, second intensity value 0.03.

[0053] The current S4 calculation yields a first intensity value of 0.07 and a second intensity value of 0.065. Comparing the current value with the corresponding values ​​in the historical records: 0.07 is more than twice the historical maximum first intensity value of 0.031 (0.07 > 0.062), therefore the current acceleration phase is determined to be an abnormal acceleration. Similarly, 0.065 is also more than twice 0.03 (0.065 > 0.06), so the deceleration phase is also determined to be an abnormal deceleration. Next, the duration of the acceleration phase in the current trend feature sequence is extracted. The acceleration phase consists of six consecutive time points P1 to P6, with each time point spaced five seconds apart, and the duration is measured by the number of time points, which is 6. The preset length threshold is 4. Since 6 is greater than 4, the condition is met, and a reclassification request is initiated.

[0054] The reclassification request is sent to a dedicated reclassification decision maker. The decision maker performs the following operations: it sorts all parameter types in the set of parameters to be revoked according to their average sensitivity scores from highest to lowest. In this embodiment, there are five groups with scores of: convolutional layer weights (95), batch normalized scaling (82), batch normalized offset (76), fully connected layer weights (55), and output layer bias (20). After sorting, the top 20% (20% of the five groups) are forcibly assigned to the high-sensitivity parameter group; that is, the convolutional layer weights are forcibly assigned to the high-sensitivity parameter group and configured with an instantaneous zeroing forgetting strategy. The bottom 20% are also forcibly assigned to the low-sensitivity parameter group; that is, the output layer bias is forcibly assigned to the low-sensitivity parameter group and configured with a linear bottom forgetting strategy. The remaining 63 groups (batch normalized scaling, batch normalized offset, and fully connected layer weights) are divided into high-sensitivity, medium-sensitivity, and low-sensitivity groups. According to the rules, the combined sensitivity weights corresponding to scores of 82 and 76 are 82 respectively. 0.07 = 5.74, 76 0.07 = 5.32, both falling within the second threshold range of 2.0~10.0, and should be classified as a high-sensitivity parameter group; the group with a score of 55 has a comprehensive weight of 3.85, also falling within the second range. However, the reclassification rule requires that if these 60% of the groups still fall into the high-sensitivity or low-sensitivity parameter groups after being classified using the original method, they should be adjusted to medium-sensitivity parameter groups and configured with an exponential decay forgetting strategy. Therefore, batch normalization scaling, batch normalization offset, and fully connected layer weight grouping were all adjusted to medium-sensitivity parameter groups and configured with an exponential decay forgetting strategy.

[0055] If the current intensity value does not exceed 1.5 times the historical value (e.g., the current first intensity value is 0.04, the historical maximum is 0.031, and 0.04 < 0.0465), it is considered a stationary change and is directly classified without reclassification. If only the first intensity value exceeds 1.5 times the historical value but does not reach 2 times (e.g., 0.045), and the second intensity value does not exceed 1.5 times, it is considered a unilateral anomaly. In this case, the parameter type grouping related to the acceleration-side change is identified, such as the convolutional layer weight grouping and batch normalized scaling grouping directly related to heart rate acceleration. The average sensitivity score of these two groups is multiplied by the enhancement coefficient of 1.8 to obtain the corrected score, and then the comprehensive sensitivity weight is recalculated based on the corrected score before grouping. Other groups, such as output layer bias, retain their original comprehensive sensitivity weights.

[0056] Furthermore, the health monitoring event stream corresponding to the health digital identity in S1 is serialized event data received in real time from the wearable device interface, and the extraction process of the trend feature sequence is specifically as follows: For each event in the health monitoring event stream, extract the event occurrence timestamp and raw health indicator readings, and arrange the raw health indicator readings into a raw reading sequence. In the original reading sequence, a sampling point is taken every fixed number of events to obtain a sampling point sequence. The difference between adjacent sampling points in the sampling point sequence is calculated, and the difference is divided by the time interval between adjacent sampling points to obtain the rate of change sequence. The rate of change sequence is filtered by moving average, with the window size of the moving average set to three rate of change values ​​to obtain a smoothed rate of change sequence. Then, three consecutive smoothed rate of change values ​​in the smoothed rate of change sequence are combined into a trend feature vector. All trend feature vectors are arranged in chronological order to form a trend feature sequence.

[0057] In this embodiment, a connection is established with a smartwatch via Bluetooth Low Energy. The smartwatch collects three health indicators from the user every two seconds: skin conductance response (SFR), heart rate, and cadence, and encapsulates these three values ​​into an event. The event format is as follows: the event header contains an event type code (0x11), an event length byte, a four-byte unsigned integer timestamp in seconds (based on January 1, 2000), a two-byte signed integer for the SFR value, a one-byte unsigned integer for the heart rate value, a one-byte unsigned integer for the cadence value, and a tail checksum. The event receiving thread continuously listens to the Bluetooth port. Whenever a complete event packet is received, it is placed into an unlocked circular buffer with a capacity of over 1020 events. When the buffer is full, the oldest event is discarded, and the discard count is recorded.

[0058] The extraction process of the trend feature sequence is triggered periodically by a separate extraction thread. Each trigger processes the most recent 256 events. The extraction thread first reads all events from the circular buffer in chronological order of timestamps, and extracts the timestamp and raw heart rate readings from each event. This embodiment uses the heart rate index as an example for illustration. The resulting raw reading sequence R has a length of L. For example, R[0]=72, R[1]=73, R[2]=74, R[3]=72, R[4]=75, R[5]=77, …… Next, a sampling point is taken from the original reading sequence at fixed intervals. The fixed interval is set to one sampling point every two events, that is, one sampling point is taken every other event. Starting from the first event, events with indices 0, 2, 4, 6, ... are taken, thus obtaining a sampling point sequence S with a length of approximately L / 2. For example, S[0]=72, S[1]=74, S[2]=75, ... Then calculate the difference between adjacent sampling points in the sampling point sequence: D[i] = S[i+1] - S[i]. Then divide the difference D[i] by the time interval between adjacent sampling points. Since the sampling interval is the time interval between every two events, and each event is two seconds apart, the time interval between two sampling points is four seconds. The rate of change C[i] = D[i] / 4. For example, S[0]=72, S[1]=74, D=2, C=0.5; S[1]=74, S[2]=75, D=1, C=0.25.

[0059] After obtaining the rate of change sequence C, a moving average filter is applied. The window size of the moving average is set to three rate of change values. For the rate of change sequence C[0], C[1], C[2], C[3], C[4], …, the smoothed rate of change sequence L is calculated: L[0] = (C[0] + C[1] + C[2]) / 3, L[1] = (C[1] + C[2] + C[3]) / 3, L[2] = (C[2] + C[3] + C[4]) / 3, and so on. Each L[i] retains six decimal places.

[0060] Finally, three consecutive smoothed rate of change values ​​in the smoothed rate of change sequence are combined into a trend feature vector. For example, L[0], L[1], L[2] are taken to form the first trend feature vector V0=(L[0], L[1], L[2]); L[1], L[2], L[3] are taken to form the second trend feature vector V1 = (L[1], L[2], L[3]); and so on. All trend feature vectors are arranged in chronological order to form a trend feature sequence. The first component of each trend feature vector in this sequence is the first smoothed rate of change, the second component is the second smoothed rate of change, and the third component is the third smoothed rate of change. The time point corresponding to the second smoothed rate of change is located between the time points corresponding to the first smoothed rate of change and the third smoothed rate of change. Because the moving average window slides, the weight of the middle point is the highest. This trend feature sequence will be used as the input data in the subsequent S2.

[0061] According to a second embodiment of the present invention, the present invention claims protection for a federated learning dynamic revocation and parameter-level forgetting system based on health digital identity, comprising: One or more processors; A memory storing one or more programs that, when executed by one or more processors, enable the processors to implement the described method for dynamic revocation and parameter-level forgetting in federated learning based on health digital identity.

[0062] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A federated learning dynamic revocation and parameter-level forgetting method based on health digital identity, characterized in that, Includes the following steps: S1: Collect the health monitoring event stream corresponding to the health digital identity, and extract the trend feature sequence of the health digital identity from the health monitoring event stream; S2: Based on the continuous change slope and fluctuation amplitude in the trend feature sequence, determine whether the triggering condition for dynamic cancellation is met within the current time window. If it is met, generate a cancellation start command. S3: In response to the revocation initiation command, locate all participating nodes associated with the health digital identity in the federated learning architecture, and parse out the set of parameters to be revoked; S4: Calculate the degree of change for the accelerated and decelerated change segments in the trend feature sequence, group each parameter type in the set of parameters to be undone, and configure the corresponding forgetting strategy. S5: Perform parameter-level forgetting operation on each parameter in the sensitive parameter group according to the forgetting strategy; S6: After completing the parameter-level forgetting, write the timestamp of this revocation and the type marker of the forgotten parameters into the digital identity cancellation record table corresponding to the health digital identity.

2. The method according to claim 1, characterized in that, The triggering condition for dynamic cancellation in S2 is determined in the following way: Extract health indicator values ​​for three consecutive time points from the trend feature sequence, calculate the first slope of change from the first time point to the second time point, and the second slope of change from the second time point to the third time point. If both the first slope of change and the second slope of change are greater than a preset positive threshold and the difference between the first slope of change and the second slope of change is less than a preset stable deviation, then it is determined to be a continuous positive acceleration trend and dynamic cancellation is triggered. If both the first change slope and the second change slope are less than a preset negative threshold and the difference between the first change slope and the second change slope is less than a preset stable deviation, then it is determined to be a continuous negative acceleration trend and dynamic cancellation is triggered. If the product of the first slope and the second slope is negative and the fluctuation amplitude of the trend feature sequence exceeds the preset violent fluctuation threshold in the three consecutive time points, it is determined to be a trend reversal fluctuation and dynamic cancellation is triggered.

3. The method according to claim 2, characterized in that, S3 includes: Read the identification code of the health digital identity, input it into the pre-established node mapping directory, and obtain the list of first participating nodes through hash lookup; Then, based on the node association frequency extracted from the historical interaction records of the health digital identity, cold nodes with association frequencies lower than a preset frequency threshold are filtered out from the first participating node list to obtain the second participating node list; For each participating node in the second participating node list, a parameter query request is sent to that participating node, the parameter index table returned by the participating node is received, and the parameter storage address marked with the health digital identity tag is selected from the parameter index table to form the set of parameters to be revoked.

4. The method according to claim 1, characterized in that, S4 performs the following operations: The trend feature sequence is segmented and identified. A sliding time window is set, and the rate of change between each pair of adjacent time points within the sliding time window is calculated to obtain multiple rate of change values. If the rate of change values ​​increases sequentially and the difference between adjacent rate of change values ​​is greater than the first increment threshold, then the sliding time window is marked as an accelerated change segment. If these multiple rate-of-change values ​​decrease sequentially and the difference between adjacent rate-of-change values ​​is greater than the first decreasing threshold, then the sliding time window is marked as a deceleration segment. For each identified acceleration segment, the difference between the maximum and minimum rate of change within that segment is taken as the first severity value of the acceleration segment. For each identified deceleration segment, the difference between the absolute value of the maximum and the absolute value of the minimum rate of change within that segment is taken as the second severity value of the deceleration segment. Iterate through each parameter type group in the set of parameters to be revoked, which contains multiple parameter instances and is associated with parameter sensitivity labels. Add up the parameter sensitivity label values ​​of all parameter instances in the parameter type group and divide by the total number of instances in the parameter type group to obtain the average sensitivity score of the parameter type group. Multiply the average sensitivity score by the first severity value to obtain the sensitivity weight under accelerated change, and multiply the average sensitivity score by the second severity value to obtain the sensitivity weight under decelerated change. Take the larger value between the sensitivity weight under accelerated change and the sensitivity weight under decelerated change as the comprehensive sensitivity weight for this parameter type group.

5. The method according to claim 4, characterized in that, S5 includes: Iterate through all parameter storage addresses in the high-sensitivity parameter group. For each storage address, read the current parameter value and determine whether the current parameter value is equal to the preset initial parameter value. If not, set all binary bits in the storage address to logic 0 and update the checksum of the storage address to the checksum corresponding to all 0s. For each parameter in the sensitive parameter group, obtain the current value of the parameter, read the attenuation count value associated with the parameter. The attenuation count value is initially zero. If the attenuation count value is less than the maximum attenuation count, divide the current value of the parameter by the attenuation base to obtain an intermediate value. Round the intermediate value according to the data type precision of the parameter to obtain an updated value. Write the updated value back to the storage address of the parameter and increment the attenuation count value by one. If the attenuation count value has reached the maximum attenuation count, skip further attenuation operations for the parameter and mark it as completely forgotten in the parameter's metadata.

6. The method according to claim 1, characterized in that, In step S4, the degree of change is calculated for the acceleration and deceleration segments in the trend feature sequence. Based on the degree of change, each parameter type in the set of parameters to be withdrawn is grouped into a high-sensitivity parameter group, a medium-sensitivity parameter group, and a low-sensitivity parameter group. An instantaneous zeroing forgetting strategy is configured for the high-sensitivity parameter group, an exponential decay forgetting strategy is configured for the medium-sensitivity parameter group, and a linear bottoming forgetting strategy is configured for the low-sensitivity parameter group.

7. The method according to claim 1, characterized in that, Before configuring the forgetting strategy, a dynamic reclassification operation is performed for the following parameter types: Obtain the historical revocation record table of the health digital identity, which records the trend feature sequence fragments of the five most recent revocation events and the forgetting strategy identifier configured for each parameter type group at that time; The first intensity value of the acceleration change segment and the second intensity value of the deceleration change segment calculated in the current S4 are compared with the historical intensity values ​​of the corresponding trend feature sequence segments in the historical cancellation record table. If the current first intensity value exceeds twice any historical intensity value, the current acceleration change segment is determined to be an abnormal acceleration. If the current second intensity value exceeds twice any historical intensity value, the current deceleration change segment is determined to be an abnormal deceleration. When it is determined to be an abnormal acceleration or deceleration, the duration of the acceleration or deceleration segment in the current trend feature sequence is extracted. The duration is measured by the number of time points. If the duration is greater than a preset length threshold, a reclassification request is initiated. The reclassification request carries the identifier of the current trend feature sequence. In response to the reclassification request, all parameter type groups in the set of parameters to be revoked are sorted from high to low according to their average sensitivity scores. The top 20% of the parameter type groups are forcibly assigned to the high-sensitivity parameter group and configured with an instantaneous zeroing forgetting strategy. The bottom 20% of the parameter type groups are forcibly assigned to the low-sensitivity parameter group and configured with a linear bottoming forgetting strategy. The remaining parameter type groups are divided into high-sensitivity, medium-sensitivity, and low-sensitivity parameter groups and configured with corresponding strategies.

8. The method according to claim 1, characterized in that, The health monitoring event stream corresponding to the health digital identity in S1 is serialized event data received in real time from the wearable device interface. The extraction process of the trend feature sequence is as follows: For each event in the health monitoring event stream, extract the event occurrence timestamp and raw health indicator readings, and arrange the raw health indicator readings into a raw reading sequence. In the original reading sequence, a sampling point is taken every fixed number of events to obtain a sampling point sequence. The difference between adjacent sampling points in the sampling point sequence is calculated, and the difference is divided by the time interval between adjacent sampling points to obtain the rate of change sequence. The rate of change sequence is filtered by moving average, with the window size of the moving average set to three rate of change values ​​to obtain a smoothed rate of change sequence. Then, three consecutive smoothed rate of change values ​​in the smoothed rate of change sequence are combined into a trend feature vector. All trend feature vectors are arranged in chronological order to form a trend feature sequence.

9. A federated learning dynamic revocation and parameter-level forgetting system based on health digital identity, characterized in that, include: One or more processors; A memory having stored one or more programs that, when executed by one or more processors, cause the one or more processors to implement a federated learning dynamic revocation and parameter-level forgetting method based on health digital identity as described in any one of claims 1 to 8.

Images