Federal learning worker selection method based on multi-dimensional constraint and storage medium

Abstract

The invention discloses a federated learning worker selection method based on multi-dimensional constraints and a storage medium, and the method comprises the steps that a cloud platform sets a multi-dimensional resource constraint condition and constructs a constraint region according to a federated learning task demand; collecting multi-dimensional attribute information of candidate workers, forming an attribute vector, calculating a comprehensive evaluation value, and carrying out priority ranking on the workers according to the evaluation value to realize preliminary selection; and for all the workers, calculating Euclidean distances between the attribute vectors of the workers and the optimal point and the most secondary point in the boundary of the constraint region, and complementarily selecting a plurality of workers closest to the optimal point and farthest from the most secondary point so as to realize boundary optimization and resource matching. According to the method, multi-dimensional attribute evaluation is combined with spatial distance analysis, the task bottleneck caused by resource limitation is remarkably relieved in the heterogeneous Internet of Things environment, the synchronization waiting time in the federated learning aggregation process is shortened, and the task completion efficiency and the system robustness are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and Internet of Things technology, and relates to federated learning and mobile crowd sensing, and particularly to a worker selection method and system based on multidimensional resource constraints suitable for heterogeneous Internet of Things environments. Background Technology

[0002] In the mobile crowd sensing scenario of federated learning, participating devices generally have the following characteristics: significant differences in computing and storage capabilities (performance heterogeneity), limited resources, and unstable network connections. These conditions pose challenges to the effective selection of workers (participating devices). Existing worker selection strategies mainly have the following shortcomings: (1) Single evaluation index. Most methods rely on a single performance parameter or simple screening rules, failing to comprehensively consider the overall performance of devices in terms of multi-dimensional resource attributes such as computing power, storage space, network bandwidth, and energy consumption. This results in a lack of balance between computing, communication, and energy consumption in the selection results, thereby reducing task execution efficiency and system stability. (2) Lack of optimized utilization of resource constraint space. Current methods usually do not systematically analyze and optimize resource constraints, lacking effective strategies to identify and utilize devices at the performance boundary or potential advantage position. This results in some devices with high task completion potential not being selected, causing waste of available resources and prolonging model aggregation time. Therefore, there is an urgent need for a worker selection method that can achieve comprehensive evaluation of device capabilities under multi-dimensional constraints and improve resource utilization, so as to improve the completion efficiency and system robustness of federated learning tasks in heterogeneous Internet of Things environments. Summary of the Invention

[0003] This invention provides a federated learning worker selection method based on multidimensional constraints to address the problems of low worker selection efficiency, resource matching imbalance, and excessively long model aggregation waiting time in heterogeneous IoT environments due to factors such as differences in device performance, limited resources, and network instability.

[0004] The technical solution of the present invention includes the following steps:

[0005] (1) Multidimensional constraint modeling and worker attribute collection:

[0006] Based on the current federated learning task, the federated mobile crowd sensing platform... Resource requirements, establish 3D constraint set ;in: . No. 3D constraints are represented as closed intervals: , used to limit the The allowed scope of resource attributes. These resource attributes include local data size. Equipment computing power Network upload bandwidth and remaining battery power The platform will handle the tasks. and constraint set The broadcast was sent to all registered workers.

[0007] (2) Quantification and constraint screening of worker capabilities:

[0008] Each worker Organize its current resource status as 3D attribute vector:

[0009] The responses are then sent to the platform; after receiving all responses, the platform creates a set: For each worker Calculate the comprehensive ability evaluation value:

[0010] The platform then performs constraint filtering: if for any All meet If so, the worker will be included in the candidate set: .

[0011] (3) Supplementary selection based on evaluation ranking and boundary distance:

[0012] candidate set according to The values ​​are sorted in descending order. Let the number of workers required for the task be... .

[0013] like Select the previous one directly The renowned workers served as the training group for this round. .

[0014] like ,but:

[0015] <1> Define the vector of the upper bound point in the constraint space: .

[0016] <2> For each of the remaining workers Calculate its Euclidean distance from the boundary points:

[0017] .

[0018] <3> Select by distance in ascending order The workers made up for it. .

[0019] (4) Collaborative privacy protection training:

[0020] Platform to Training tasks are distributed, and each worker updates the model locally. Then, an encrypted upload is performed using a distributed encryption protocol based on EC-ElGamal.

[0021] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described federated learning worker selection method based on multidimensional constraints.

[0022] The beneficial effects of this invention are:

[0023] This invention enables comprehensive evaluation and optimized selection of multi-dimensional resource attributes in heterogeneous IoT environments, balancing computing, communication, and energy consumption; combined with boundary replenishment strategies, it can fully utilize potential high-performance devices and reduce waste of available resources; and a privacy-preserving encryption mechanism ensures the data security of model updates in a distributed environment. Detailed Implementation

[0024] The technical solution of the present invention will be further described in detail below with reference to embodiments. These embodiments are illustrated in detail below.

[0025] A typical heterogeneous IoT environment is an application scenario that includes various terminal devices with different performance levels.

[0026] The system can access the federated mobile crowd sensing platform via wireless network to collaboratively complete federated learning tasks.

[0027] Example: A worker selection method based on multidimensional constraints

[0028] Step 1: Multidimensional Constraint Modeling and Task Broadcasting

[0029] During the task initialization phase, the platform determines the current federated learning training task. The resource requirements are determined, and the number of resource attribute dimensions involved is identified. (For example This may include local data scale processor

[0030] Computing power Network upload bandwidth Remaining battery power For each dimension attribute, the platform sets a constraint range: , .

[0031] The above parameters can be dynamically set according to the task type, data size, or model complexity. For example, when the model...

[0032] For larger applications requiring high computing power, the capacity can be increased. and reduce To ensure upload speed, the platform will impose restrictions on the task description (including the number of training epochs, initial model parameters, and data processing requirements) and the set of constraints. The broadcast is sent to all registered worker nodes.

[0033] Step 2: Collecting Worker Attributes and Quantifying Abilities

[0034] Each worker After receiving the task broadcast, collect the current state to form an attribute vector: .For example: Indicates the number of local data entries. This indicates the processor's floating-point computing power (FLOPS). Indicates upload bandwidth (Mbps). This represents the remaining battery percentage. Workers send attribute vectors to the platform, which then generates a response set. The platform calculates a comprehensive competency score for each worker:

[0035] This indicator, through the sum of squares calculation, can amplify the differences in advantages across dimensions, making high-performing workers stand out more in the ranking.

[0036] Step 3: Constraint Filtering and Priority Queue Construction

[0037] The platform checks each worker Does it meet the following requirements: ;

[0038] If all dimensional constraints are met, then add the worker to the candidate set: Subsequently, the candidate set was sorted according to... Sort the values ​​in descending order and build a priority queue for selection.

[0039] Step 4: Boundary Optimization and Supplementary Selection

[0040] Let the number of workers required for each round of the task be... :

[0041] like The platform directly selects the previous Workers as training .

[0042] like Execute supplementary selection:

[0043] 1. Define the vector of the upper bound point in the constraint space:

[0044] .

[0045] 2. For workers not selected in the candidate set, calculate their Euclidean distance to the upper bound:

[0046] .

[0047] 3. Sort all workers to be added in ascending order of distance, and select the one closest to the boundary point. famous workers joined .

[0048] Step 5: Distributed Training and Privacy Protection

[0049] Platform to Distribute task parameters. Each worker executes the training locally based on their own data.

[0050] Practice, Generate Model Update To protect privacy, workers employ a distributed architecture based on EC-ElGamal.

[0051] Encryption protocol encryption.

[0052] Operation process management:

[0053] The platform monitors the status of each worker during each training round. If a node is detected to be offline or has timed out, then...

[0054] Replace workers from the remaining unselected workers in turn, prioritizing those selected. High or Small nodes,

[0055] To ensure the successful completion of the mission.

Claims

1. A federated learning worker selection method based on multidimensional constraints, characterized in that, The method includes the following steps: (1) Based on the current federated learning task, the federated mobile swarm intelligence sensing platform... Resource requirements, establish 3D constraint set ;in: ;No. 3D constraints are represented as closed intervals: , used to limit the The allowed scope of resource attributes; the resource attributes include local data scale. Equipment computing power Network upload bandwidth and remaining battery power The platform will handle the tasks. and constraint set The broadcast was sent to all registered workers; (2) Each worker Organize its current resource status as 3D attribute vector: The responses are then sent to the platform; after receiving all responses, the platform creates a set: For each worker Calculate the comprehensive ability evaluation value: The platform then performs constraint filtering: if for any All meet If so, the worker will be included in the candidate set: ; (3) Candidate set according to Sort the values ​​in descending order; let the number of workers required for the task be... : like Select the previous one directly The renowned workers served as the training group for this round. ; like ,but: <1> Define the vector of the upper bound point in the constraint space: ; <2> For each of the remaining workers Calculate its Euclidean distance from the boundary points: ； <3> Select by distance in ascending order The workers made up for it. ; (4) Platform to Training tasks are distributed, and each worker updates the model locally. Then, an encrypted upload is performed using a distributed encryption protocol based on EC-ElGamal.

2. The method according to claim 1, characterized in that, The method also includes a node dropout fault tolerance mechanism: a candidate worker list is established, which consists of the remaining workers who were not selected as participants in step (3); during task execution, if a selected participant is detected to be offline or has a response timeout, a candidate worker is selected from the candidate worker list based on comprehensive capability indicators. descending order or Euclidean distance In ascending order, new workers are selected to replace offline participants and join the training.

3. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method of claim 1 or 2.

Images