CN-122027621-A - Federal learning-based cross-domain video monitoring resource collaborative scheduling method and system

CN122027621ACN 122027621 ACN122027621 ACN 122027621ACN-122027621-A

Abstract

The invention relates to the technical field of federal learning, in particular to a cross-domain video monitoring resource collaborative scheduling method and system based on federal learning. Collecting original behavior data of a target node, constructing a structured behavior evidence chain, extracting the behavior evidence chain of the target node, calculating a learning reputation factor, a cooperation reputation factor and a network reputation factor of the target node, calculating reputation confidence of the target node in a preset evaluation period, judging the stage of the system based on the operation data, distributing dynamic weights for the learning reputation factor, the cooperation reputation factor and the network reputation factor based on the stage, weighting the learning reputation factor, the cooperation reputation factor and the network reputation factor based on the dynamic weights, synthesizing a comprehensive reputation value of the target node, and jointly forming a reputation state vector of the target node by the comprehensive reputation value and the reputation confidence, and executing a differential scheduling strategy for the target node according to the reputation state vector of the target node in resource cooperative scheduling.

Inventors

ZHU SHENGJUN
WU ZHIWEI
Jia Wukuo
WANG HONGBANG

Assignees

雄安零度科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260302

Claims (10)

1. The federal learning-based cross-domain video monitoring resource collaborative scheduling method is characterized by comprising the following steps of: In a preset evaluation period, acquiring original behavior data of a target node, and constructing a structured behavior evidence chain, wherein the original behavior data is composed of data of three dimensions of federal learning behavior, task cooperation behavior and network communication behavior; Extracting a behavior evidence chain of the target node, and calculating a learning reputation factor, a cooperation reputation factor and a network reputation factor of the target node based on a preset regularization algorithm; Evaluating the behavior evidence chain based on a preset rule, obtaining an integrity score, a cross-source consistency score and an evidence source authority score, and calculating the credit confidence of the target node in a preset evaluation period; Judging a stage in which the system is positioned based on the operation data, and distributing dynamic weights for learning reputation factors, collaborative reputation factors and network reputation factors based on the stage, wherein the stage comprises a training intensive stage, a task intensive stage, a network unstable stage and a stable operation stage; Weighting the learning reputation factor, the cooperative reputation factor and the network reputation factor based on the dynamic weight, synthesizing a comprehensive reputation value of the target node, and forming a reputation state vector of the target node by the comprehensive reputation value and the reputation confidence together; In the resource cooperative scheduling, a differential scheduling strategy is executed for the target node according to the reputation state vector of the target node, wherein a priority strategy is adopted for the target node with the comprehensive reputation value and the reputation confidence reaching the preset standard, and a limiting strategy is adopted for the target node with the reputation confidence value not reaching the preset standard.
2. The federal learning-based cross-domain video monitoring resource collaborative scheduling method according to claim 1, wherein the building a structured behavioral evidence chain comprises: Configuring a unique transaction ID for each behavior record of the target node; and linking the behavior record generated by the target node, the associated task issuer record and the third party witness record through the transaction ID and aligning the time stamps to form a chain type data structure serving as the behavior evidence chain.
3. The federal learning-based cross-domain video monitoring resource collaborative scheduling method according to claim 1, wherein calculating a learning reputation factor of a target node comprises: Acquiring federal learning behaviors of the target node based on the behavior evidence chain; Extracting model update data of a target node in the federation learning of the present round from the federation learning behavior to serve as a target model update set; continuously traversing federal learning behaviors of all nodes to obtain a plurality of model update sets of a plurality of nodes; calculating the performance gain of the model update of the target node relative to the global model of the previous round by using the shadow test set; Calculating a plurality of cosine similarities between a target model updating set and the plurality of model updating sets, and taking the average value of the plurality of cosine similarities as a similarity index of the target model updating set; and respectively normalizing the similarity index and the performance gain, and then carrying out weighted summation to obtain a learning reputation factor.
4. The federal learning-based cross-domain video monitoring resource collaborative scheduling method according to claim 1, wherein calculating the collaborative reputation factor of the target node comprises: when the target node claims to complete a certain cooperative task, extracting a behavior record of the target node, and a third party witness record of a plurality of neighborhood nodes of the target node about the cooperative task in the same space-time window; performing logic consistency comparison on the behavior record provided by the target node and the third party witness record based on a preset rule to obtain a consistency index; When the consistency index meets a preset consistency threshold, judging that the task cooperation behavior is true, otherwise, judging that the task cooperation behavior is not true; And calculating the proportion of the task cooperation behavior of the target node in the history time, and taking the value of the proportion as the cooperation reputation factor of the target node.
5. The federal learning-based cross-domain video monitoring resource co-scheduling method according to claim 1, wherein calculating the network reputation factor of the target node comprises: When a preset evaluation period starts, calculating to obtain a reference network round trip delay and a reference packet loss rate of the evaluation period; Acquiring average round trip delay of a target node when executing an actual cooperative task in a preset evaluation period; calculating and obtaining a delay score of a target node based on the average round-trip delay and the reference network round-trip delay; acquiring an average data packet loss rate of a target node in a preset evaluation period, and configuring a packet loss penalty item based on the average data packet loss rate and a preset packet loss rate tolerance threshold; acquiring a historical connection log, calculating the connection success rate of a target node in the latest N evaluation periods based on the historical connection log, and taking a sliding average value of the connection success rate as a communication stability index; And subtracting the packet loss penalty term after weighting and fusing the time delay score and the communication stability index to obtain the network reputation factor.
6. The federal learning-based cross-domain video monitoring resource collaborative scheduling method according to claim 1, wherein evaluating the behavioral evidence chain based on a preset rule, obtaining an integrity score, a cross-source consistency score, and an evidence source authority score, and calculating a reputation confidence of a target node in a preset evaluation period, comprises: acquiring a preset necessary evidence item of a behavior evidence chain, searching the behavior evidence chain, acquiring the deletion proportion of the necessary evidence item, and taking the deletion proportion as the integrity score; Extracting evidence of different sources in a behavior evidence chain, calculating the consistency degree of the evidence of the different sources on key assertions, and taking the consistency degree as the cross-source consistency score; Extracting a plurality of reputation confidences of a plurality of nodes providing the third party witness records, carrying out mean value calculation on the reputation confidences, and taking the mean value as the evidence source authority score; And carrying out weighted fusion on the integrity score, the cross-source consistency score and the evidence source authority score to obtain the credit confidence.
7. The federal learning-based cross-domain video monitoring resource collaborative scheduling method according to claim 1, wherein the stage in which the system is located is determined based on operational data and dynamic weights are assigned to the learning reputation factor, the collaborative reputation factor, and the network reputation factor based on the stage in which the system is located, wherein the stage includes a training intensive stage, a task intensive stage, a network instability stage, and a stationary operation stage, comprising: Counting the number of concurrently performed federal learning training tasks in a system, and judging as a training intensive stage when the number of federal learning training tasks exceeds a set proportion of the total active task number and the average update amplitude of a global model exceeds a preset amplitude threshold value, and synchronously improving the learning weight of the learning reputation factor; When the number of video analysis task requests exceeds a threshold value of the number of requests in unit time, and the calculated resource utilization rate continuously exceeds a threshold value of the preset resource utilization rate in a preset time window, judging that the task is in a dense stage, and synchronously improving the cooperative weight of the cooperative reputation factor; When the variance of the average communication delay among all network nodes continuously exceeds an alarm threshold value in a preset time window, judging that the network is unstable, and synchronously improving the network weight of the network reputation factor; when the training intensive stage, the task intensive stage, or the network unstable stage is not determined, the stationary operation stage is determined, and a default weight allocation is used.
8. The federal learning-based cross-domain video monitoring resource collaborative scheduling method according to claim 7, wherein weighting the learning reputation factor, collaborative reputation factor, and network reputation factor based on the dynamic weights, synthesizing a composite reputation value of a target node, and forming a reputation state vector of the target node with the composite reputation value and the reputation confidence together comprises: The learning reputation factor, the cooperation reputation factor and the network reputation factor are weighted and summed by using the learning weight, the cooperation weight and the network weight, and the obtained summation value is multiplied by a preset historical reputation attenuation function to obtain the comprehensive reputation value; The historical reputation decay function = R (t-1) × [ γ+ (1- γ) x tanh (α×Δr) ]; wherein R (t-1) is the comprehensive credit value of the target node in the last evaluation period; gamma is a preset basic attenuation coefficient; Tanh is a hyperbolic tangent function; Δr is the trend of the target node in the change of the comprehensive reputation value in the last N continuous evaluation periods; Alpha is a trend sensitivity parameter.
9. The federal learning-based cross-domain video monitoring resource collaborative scheduling method according to claim 1, wherein performing a differentiated scheduling policy on a target node according to a reputation state vector of the target node comprises: Setting a first confidence coefficient threshold when a node is selected by a federal learning client, wherein when the credit confidence coefficient of a target node is more than or equal to the first confidence coefficient threshold, the comprehensive credit value of the target node is positively correlated with the probability that the target node is selected by the client; Setting a second confidence coefficient threshold when the federal learning model is aggregated, and only when the credit confidence coefficient of the target node is more than or equal to the second confidence coefficient threshold, incorporating model updating corresponding to the target node into the aggregation, wherein the aggregation weight of the target node is in direct proportion to the credit confidence coefficient; When monitoring task scheduling, a task publisher designates the minimum reputation requirement and the minimum confidence requirement, and a scheduler only distributes tasks to target nodes with the comprehensive reputation value more than or equal to the minimum reputation requirement and the reputation confidence more than or equal to the minimum confidence requirement.
10. The federal learning-based cross-domain video monitoring resource collaborative scheduling system is characterized in that the system is used for realizing the federal learning-based cross-domain video monitoring resource collaborative scheduling method according to any one of claims 1-9, and the system comprises: The system comprises a behavior evidence chain construction module, a target node analysis module and a network communication module, wherein the behavior evidence chain construction module is used for acquiring original behavior data of the target node in a preset evaluation period and constructing a structured behavior evidence chain, wherein the original behavior data is composed of data of three dimensions of federal learning behavior, task cooperation behavior and network communication behavior; The reputation factor calculation module is used for extracting a behavior evidence chain of the target node and calculating a learning reputation factor, a cooperation reputation factor and a network reputation factor of the target node based on a preset regularization algorithm; The credit confidence calculation module is used for evaluating the behavior evidence chain based on a preset rule, obtaining an integrity score, a cross-source consistency score and an evidence source authority score, and calculating the credit confidence of the target node in a preset evaluation period; The dynamic weight distribution module is used for judging the stage of the system based on the operation data and distributing dynamic weights for the learning reputation factors, the cooperative reputation factors and the network reputation factors based on the stage, wherein the stage comprises a training intensive stage, a task intensive stage, a network unstable stage and a stable operation stage; The comprehensive reputation value calculation module is used for weighting the learning reputation factors, the cooperative reputation factors and the network reputation factors based on the dynamic weights, synthesizing the comprehensive reputation value of the target node, and forming the reputation state vector of the target node by the comprehensive reputation value and the reputation confidence together; And the scheduling policy execution module is used for executing a differential scheduling policy on the target node according to the reputation state vector of the target node in the resource collaborative scheduling, wherein a priority policy is adopted for the target node with the comprehensive reputation value and the reputation confidence reaching the preset standard, and a limiting policy is adopted for the target node with the reputation confidence value not reaching the preset standard.

Description

Federal learning-based cross-domain video monitoring resource collaborative scheduling method and system Technical Field The invention relates to the technical field of federal learning, in particular to a cross-domain video monitoring resource collaborative scheduling method and system based on federal learning. Background In the federal learning-based cross-domain video monitoring collaborative scheduling, computing resources, data and models of all participating domains are required to be collaborative on the premise of not sharing original data. However, the prior art multi-idealized assumes that all participating nodes are in honest cooperation, and lacks an efficient recognition and governance mechanism for malicious nodes, selfish nodes, and low quality models. This results in resource scheduling imbalances, contaminated model aggregation, overall system performance and security degradation. Therefore, developing a method capable of quantifying the behavior credibility of nodes and performing scheduling and model cooperation based on the behavior credibility is a problem to be solved in the prior art. Disclosure of Invention Aiming at the technical problem that the prior art lacks an effective identifying and managing mechanism for malicious nodes, selfish nodes and low-quality models, the invention provides a federal learning-based cross-domain video monitoring resource collaborative scheduling method and system for solving the problem. The technical scheme for solving the technical problems is as follows: The invention provides a federal learning-based cross-domain video monitoring resource collaborative scheduling method, which comprises the steps of collecting original behavior data of a target node in a preset evaluation period and constructing a structured behavior evidence chain, wherein the original behavior data consists of three dimensionalities of data of federal learning behaviors, task collaboration behaviors and network communication behaviors, extracting the behavior evidence chain of the target node, calculating a learning reputation factor, a collaboration reputation factor and a network reputation factor of the target node based on a preset regularization algorithm, evaluating the behavior evidence chain based on a preset rule, obtaining an integrity score, a cross-source consistency score and an evidence source authority score, calculating reputation confidence of the target node in the preset evaluation period, judging a stage of the system based on the operation data, distributing dynamic weights for the learning reputation factor, the collaboration reputation factor and the network reputation factor based on the stage, weighting the learning factor, the collaboration reputation factor and the network reputation factor based on the preset regularization algorithm, setting a comprehensive reputation value and a reputation of the reputation node to be in a preset state, and setting the reputation of the integrated node to be different from the reputation node according to a preset state, and setting the reputation value of the integrated reputation has a preset state and a reputation value in the integrated state, and setting the integrated state and the reputation value of the reputation has a preset state and a priority value of the reputation has a priority to be different from the target node. Optionally, the construction of the structured behavioral evidence chain comprises configuring a unique transaction ID for each behavioral record of the target node, linking the behavioral record generated by the target node, the associated task publisher record and the third party witness record through the transaction IDs and aligning time stamps to form a chain type data structure serving as the behavioral evidence chain. The method comprises the steps of obtaining federal learning behaviors of a target node based on a behavior evidence chain, extracting model update data of the target node in federal learning of the target node from the federal learning behaviors to serve as a target model update set, continuously traversing federal learning behaviors of all nodes to obtain a plurality of model update sets of a plurality of nodes, calculating performance gains of model update of the target node relative to a global model of a previous round by using a shadow test set, calculating a plurality of cosine similarities of the target model update set and the plurality of model update sets, taking an average value of the cosine similarities as a similarity index of the target model update set, normalizing the similarity index and the performance gains respectively, and then weighting and summing to obtain a learning reputation factor. Optionally, calculating the cooperative reputation factor of the target node, wherein the cooperative reputation factor comprises the steps of extracting a behavior record of the target node when the target node claims to finish a certain cooperative task and a