A health crowd intelligence perception system and its cost-optimized federated learning method

Abstract

The present invention discloses a health crowd intelligence perception system and a cost-optimized federated learning method thereof. The present invention proposes a cloud-edge system architecture and a cost-optimized federated learning model, and deploys federated learning to train a global model in a cloud-edge collaborative manner; in the cost-optimized federated learning model, the edge decides whether to participate in the current round of training based on the verification accuracy of the global model to be trained, so that the communication cost and local training cost can be optimized, and the cloud performs quality assessment on the received local model updates and selects high-quality local model updates to participate in the aggregation of the global model, thereby improving learning efficiency. A large number of experiments based on public data sets have proved that the method proposed by the present invention effectively reduces communication costs and training costs while ensuring the accuracy of the global model.

Description

Technical Field

[0001] The present invention belongs to the field of health care technology, and in particular relates to a health crowd intelligence perception system and a cost-optimized federated learning method thereof. Background Art

[0002] Federated learning is a machine learning paradigm in which multiple parties hold data for collaborative training under privacy protection, and its goal is to train high-quality global models. It obtains machine learning models by iteratively updating local models and aggregating global model parameters. In each round of learning algorithm in this paradigm, each client holding data independently trains the model based on its local data and sends the model parameters to the central coordinator, who aggregates the local model parameters of the client and calculates a new global model. The learning algorithm is repeated until the global model meets certain convergence criteria. Federated learning has been successful in some applications, such as Google Gboard's next word prediction, "Hey Siri" sound classifier, and mobile augmented reality. Researchers are also actively exploring its application in medical and health diagnosis and treatment scenarios.

[0003] In the field of health care, many highly prevalent diseases (such as tinnitus, migraine, and chronic pain) lack universal and effective treatments and require long-term health monitoring. Since 2010, research on health crowd intelligence technology has made a series of progress, including HealthKit and Google Fit, which are used to create health crowd intelligence services in the Apple and Google ecosystems respectively. Due to the strong privacy characteristics of medical data, patient data is usually stored independently by hospitals or community medical centers. When establishing health monitoring and disease prediction models, the data of a single hospital or community medical center is not enough to train the model, and multiple collaborations are required to train the model to extract group intelligence.

[0004] Although federated learning provides an effective solution for extracting group intelligence in health crowd sensing (HCS), it still faces many challenges in the process of collaborative model training, such as the computational and time costs of local model training and the transmission cost of model parameters. The structure of current models is becoming increasingly complex, with millions of parameters. It is inevitable that a large amount of traffic will be consumed when transmitting parameters. Taking the VGG16 model as an example, the size of its parameters is about 57MB. If 100 hospitals or community medical centers participate in the collaborative training model and perform 10 training rounds, the model parameters will consume 57GB of transmission costs. Summary of the invention

[0005] Aiming at the problems of high model training cost and parameter transmission cost in the federated learning process in the existing health crowd intelligence perception scenario, the present invention proposes a health crowd intelligence perception system and a cost-optimized federated learning method thereof.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In one aspect, the present invention proposes a health crowd intelligence perception system, comprising: a central coordinator, an edge node and an end user; the central coordinator is a health department or a hospital consortium; the edge node is a hospital or a community medical center; the end user is an intelligent terminal carried by a patient, which is used to collect patient data and upload the patient data to the edge node;

[0008] The central coordinator and edge nodes jointly deploy federated learning, which realizes intelligent extraction with privacy protection. Edge nodes and end users collaborate in data collection and management.

[0009] Furthermore, the service process of the health crowd intelligence perception system is as follows:

[0010] Any edge node acts as the initiator of model training, selects a training model as needed, initializes the model using local data, and submits the model training requirements and the initialized model to the central coordinator;

[0011] The central coordinator distributes the global model to be trained to the edge nodes participating in the training, obtains the local model parameters of the edge nodes after training, performs quality assessment on the model parameters, selects the local model parameters participating in the global model aggregation based on the quality assessment results, and generates a new global model;

[0012] The edge node stores the data uploaded by the smart terminal, receives and verifies the global model issued by the central coordinator, and as the data holder, trains the global model with local data to generate a new local model and uploads the model parameters to the central coordinator;

[0013] Patients perceive and generate individual in-depth data through the smart terminals they carry, upload the data to edge nodes, and receive personalized diagnosis and treatment services.

[0014] Another aspect of the present invention provides a cost-optimized federated learning method in a health crowd intelligence perception system. The method is based on the health crowd intelligence perception system and includes:

[0015] Establish a cost-optimized federated learning model based on the goals of federated learning;

[0016] Before each local training, the edge node verifies the received global model with the verification data set in the real-time data. If its verification accuracy is higher than the verification accuracy of the previous round, it participates in this training, otherwise it refuses to participate in this training;

[0017] When edge nodes perform local training in each communication round, they conduct multiple epochs of training and select the model parameters with the highest verification accuracy and upload them to the central coordinator for model aggregation;

[0018] After receiving the local model parameters from the edge nodes, the central coordinator conducts quality assessment on the local model parameters using the truth value discovery method and selects the local model parameters that meet the quality assessment requirements to participate in the aggregation of the global model.

[0019] Furthermore, the establishment of a cost-optimized federated learning model according to the goal of federated learning includes:

[0020] According to the goal of federated learning,

[0021]

[0022] in f:R d →R is the global loss function, f k :R d →R is the local loss function of edge node k, θ∈R d is the optimizable model parameter, θ * is the optimized model parameter, p k is the ratio of the amount of training data of edge node k to the total amount of training data participating in the global model aggregation, l(θ,x) is the loss function of data sample x when the model is θ, {1,2,...,K} is the set of edge nodes trained by the federated learning model, k is any edge node, and D k and |D k | are the training data set and training data volume of the kth edge node respectively;

[0023] Building a cost-optimized federated learning model:

[0024]

[0025]

[0026]

[0027] st(μ-3σ) k <(μ+3σ)

[0028]

[0029]

[0030]

[0031]

[0032] Where d is the sum of the training data of the edge nodes where the local model parameters participating in the global model aggregation are located in the current training round, and d K ​is the sum of the training data of the edge nodes participating in the training, p k and are the ratio of the amount of training data of the kth edge node to the total amount of training data participating in the global model aggregation and to the total amount of overall training data, θ t represents the global model of the tth round, represents the local model of the kth edge node in the tth round, a k is the model accuracy of the kth edge node, μ and σ are the mean and standard deviation of the accuracy of all local models participating in the training, respectively.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] Starting from the unique attributes of health group intelligence perception, the present invention introduces the service paradigm of cloud-edge architecture for the first time. Aiming at the strong privacy characteristics in the health field, federated learning is deployed to extract group intelligence; the federated learning process is innovatively improved to reduce the model training cost and parameter transmission cost in the federated learning process, and effectively improve the aggregation efficiency of the global model. The performance of the method proposed in the present invention is compared and analyzed with the FedAvg algorithm, which proves that the method of the present invention reduces the training cost while reducing the communication cost. Through a large number of experiments based on open data sets, it is shown that the communication cost, local model training time and computational cost of the method of the present invention are superior to the FedAvg algorithm, which verifies the effectiveness of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic diagram of the architecture of a health crowd intelligence perception system according to an embodiment of the present invention;

[0036] Figure 2 This is a service flow chart of a health crowd intelligence perception system according to an embodiment of the present invention;

[0037] Figure 3 This is a comparison chart of the local model training duration for each round in an embodiment of the present invention;

[0038] Figure 4 The embodiment of the present invention randomly selects 4 nodes to participate in the local model training;

[0039] Figure 5 A comparison diagram of communication cost and computation cost in different algorithms of the embodiments of the present invention;

[0040] Figure 6 This is a comparison chart of the global model accuracy and training loss in different algorithms of the embodiments of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0042] In order to track the health status of patients and obtain individual in-depth data for common disease research and personalized medical guidance for patients, the Health Crowd Intelligence Perception System (HCS) collects in-depth data related to individual health through a large number of terminals such as smartphones and wearable devices carried by patients. These data are stored in hospitals or community medical centers in the form of isolated islands. More importantly, the group intelligence extracted from these data will be used for the research of common diseases and personalized medical guidance for patients.

[0043] In order to comprehensively learn patient data without compromising privacy and security, the present invention proposes a health crowd intelligence perception system, which is a cloud-edge-based HCS architecture. Through federated learning, a comprehensive understanding of global data can be achieved to solve the data island problem in intelligent extraction (global model training).

[0044] like Figure 1 As shown, a health crowd intelligence perception system includes: a central coordinator (cloud), an edge node (edge) and an end user (end); the central coordinator is a health department or a hospital consortium; the edge node is a hospital or a community medical center; the end user is an intelligent terminal (such as a smart phone, a wearable device) carried by the patient, which is used to collect patient data and upload the patient data to the edge node;

[0045] The central coordinator and edge nodes jointly deploy federated learning, which realizes intelligent extraction with privacy protection. Edge nodes and end users collaborate in data collection and management.

[0046] Furthermore, if Figure 2 As shown, the service process of the health crowd intelligence perception system is:

[0047] 1) Any edge node acts as the initiator of model training, selects a training model as needed, initializes the model using local data, and submits the model training requirements and the initialized model to the central coordinator;

[0048] 2) The central coordinator distributes the global model to be trained to the edge nodes participating in the training, obtains the local model parameters of the edge nodes after training, performs quality assessment on the model parameters, selects the local model parameters participating in the global model aggregation according to the quality assessment results, and generates a new global model;

[0049] 3) The edge node stores the data uploaded by the smart terminal, receives and verifies the global model issued by the central coordinator, and as the data holder, trains the global model with local data to generate a new local model and uploads the model parameters to the central coordinator;

[0050] 4) Patients perceive and generate individual in-depth data through the smart terminals they carry, upload the data to edge nodes, and receive personalized diagnosis and treatment services.

[0051] Specifically, when a patient joins HCS, he or she needs to complete registration at the hospital or community treatment center to which he or she belongs and obtain a unique user ID. After that, he or she can complete data collection and transmission according to the prompts to receive personalized diagnosis and treatment services provided by the hospital or community treatment center. Since the present invention focuses on optimizing the communication cost and training cost of federated learning, it is assumed that the patient and the hospital or community treatment center to which he or she belongs are in a secure network and that the patient provides high-quality individual in-depth health data.

[0052] In order to solve the problem that the hospitals or community treatment centers do not have enough data to train the required models, they join the HCS system as edge nodes. When joining for the first time, they need to register and bind their real identities with the central coordinator. After that, each authorized node obtains its own public / private key pair (PKen, SKen) for information encryption when communicating with the central coordinator (such as training-related information).

[0053] On the basis of the above embodiments, in order to reduce the communication cost while ensuring the global model aggregation efficiency and reducing the training cost of the local model, the present invention also proposes a cost-optimized federated learning method in a health crowd intelligence perception system, including:

[0054] 1) Establish a cost-optimized federated learning model (COFL) based on the goals of federated learning;

[0055] 2) Before each local training, the edge node verifies the received global model with the verification data set in the real-time data. If its verification accuracy is higher than the verification accuracy of the previous round, it participates in this training. Otherwise, it refuses to participate in this training (let the accuracy of this round of model a k =0,k∈[1,K], K is the number of edge nodes); this step effectively reduces the computing cost of edge devices and the communication cost between edge and cloud, and accelerates the aggregation of the model.

[0056] 3) When the edge node performs local training in each communication round, it conducts multiple epochs of training and selects the model parameters with the highest verification accuracy to upload to the central coordinator for model aggregation; this step can speed up the aggregation of the model to reduce the number of communication rounds, thereby reducing the communication cost.

[0057] 4) After receiving the local model parameters from the edge nodes, the central coordinator uses the truth value discovery method to evaluate the quality of the local model parameters and selects the local model parameters that meet the quality evaluation requirements to participate in the aggregation of the global model. This step can speed up the aggregation of the model.

[0058] Specifically, unlike the Federated Averaging Algorithm (FedAvg), which randomly selects a portion of data holders to participate in model training at the beginning of each round of training, COFL maintains the relative stability of edge nodes participating in global model training based on the inherent characteristics of healthy crowd intelligence perception, that is, all edge nodes participate in training. Once the federated learning process begins, the participating edge nodes will no longer change.

[0059] Furthermore, the establishment of a cost-optimized federated learning model as a global model according to the goal of federated learning includes:

[0060] According to the goal of federated learning,

[0061]

[0062] in f:R d →R is the global loss function, f k :R d →R is the local loss function of edge node k, θ∈R d is the optimizable model parameter, θ * is the optimized model parameter, p k is the ratio of the amount of training data of edge node k to the total amount of training data participating in the global model aggregation, l(θ,x) is the loss function of data sample x when the model is θ, {1,2,...,K} is the set of edge nodes trained by the federated learning model, k is any edge node, and D k and |D k | are the training data set and training data volume of the kth edge node respectively;

[0063] Establish a cost-optimized federated learning model (referred to as the COFL model):

[0064]

[0065]

[0066]

[0067] st(μ-3σ) k <(μ+3σ)

[0068]

[0069]

[0070]

[0071]

[0072] ​Where d is the sum of the training data of the edge nodes where the local model parameters participating in the global model aggregation are located in the current training round, and d K is the sum of the training data of the edge nodes participating in the training, p k and are the ratio of the amount of training data of the kth edge node to the total amount of training data participating in the global model aggregation and to the total amount of overall training data, θ t represents the global model of the tth round, represents the local model of the kth edge node in the tth round, a k is the model accuracy of the kth edge node, μ and σ are the mean and standard deviation of the accuracy of all local models participating in the training, respectively.

[0073] Obviously, according to Chebyshev's theorem: of all data, at least 8 / 9 (or 88.9%) of the data are within 3 standard deviations of the mean. Therefore, the COFL model selects 88.9% of the local model parameters in each round of training to participate in the aggregation of the global model, filtering out local training results with too much deviation to accelerate the convergence of the global model. From formulas (3)-(5), it can be seen that when deriving the mean and standard deviation, the COFL model fully considers the impact of different edge nodes having different amounts of data.

[0074] (1) Convergence of the COFL model

[0075] This section theoretically analyzes the convergence of the COFL model.

[0076]

[0077] As is well known, whether it is a convex loss function or a non-convex loss function, gradient descent will converge to a local minimum or saddle point, such as linear regression and neural networks. Therefore, the present invention uses the norm of the expected gradient as the convergence indicator, that is:

[0078]

[0079] The following assumptions are used:

[0080] Assumption 1: The loss functions are both f(·),f k (·) are L-smooth, that is, they are Lipschitz continuous. For f(·), f k For any x and y of (·), we have

[0081]

[0082] Assumption 2 The expected square norm of the stochastic gradient is uniformly bounded, that is,

[0083] According to the assumptions, the convergence theorem of the COFL model is as follows,

[0084] Theorem 1: Under assumptions 1 and 2, when the learning rate is set to The convergence of the model satisfies

[0085]

[0086] prove:

[0087] To derive First derive Ef(θ t+1 )

[0088] Since f(·) is an L-smooth function, its second-order Taylor expansion is

[0089]

[0090] According to the Cauchy inequality, inequality (7) can be further derived as

[0091]

[0092] Inequality (8) can be rewritten as

[0093]

[0094] Putting inequality (9) into the expected square norm of the stochastic gradient, we have

[0095]

[0096] The learning rate Substituting into inequality (10), we obtain the result of (6).

[0097]

[0098] (2) COFL model performance analysis

[0099] This section will compare and analyze the performance of COFL and FedAvg, proving that COFL can reduce both communication cost and training cost. c Global model cost issued by the central coordinator and local model costs uploaded to edge nodes constitute.

[0100]

[0101]

[0102]

[0103]

[0104] where θ t It is the global model to be trained sent by the edge nodes from the central coordination direction in the tth training round. They are the local model parameters, local model accuracy, and core training information (such as training duration and training data volume) transmitted by the kth edge node to the central coordinator in the tth training round. If the kth edge node does not participate in the training in the tth training round, the edge node only transmits To the central coordinator, there is no local model upload cost (so the communication cost of this node is ignored). Compared with models with millions of parameters, this is almost negligible, so equation (13) is adjusted to

[0105]

[0106]

[0107] The communication cost of the FedAvg algorithm is cCom a ,

[0108]

[0109]

[0110]

[0111] The training cost of federated learning consists of two parts: computing cost and time cost. Since the computing cost of aggregation by the central coordinator is significantly different from the computing cost of model training at the edge nodes, the computing cost of aggregation is ignored and only the computing cost of training is counted.

[0112] The computational costs of COFL and FedAvg algorithms are cCpu respectively. c and cCpu a ,

[0113]

[0114]

[0115]

[0116] in is the model training cost of the kth edge node in the tth training round.

[0117] Similarly, the time cost of COFL and FedAvg algorithms is cTimec and cTime a ,

[0118]

[0119]

[0120]

[0121] Next, we will give the theorem that COFL saves communication and training costs and prove it.

[0122] Theorem 2. When learning the same model and the edge nodes participating in the training are the same, the communication cost of COFL is less than or equal to the communication cost of the FedAvg algorithm, that is, cCom c ≤cCom a .

[0123] prove:

[0124] From formulas (11)-(16), we can see that

[0125]

[0126] Obviously, in the entire T-round model interaction of federated learning, as long as any edge node does not perform local training, the communication cost of COFL is better than that of the FedAvg algorithm.

[0127] Theorem 3. When learning the same model and the edge nodes participating in the training are the same, the computational cost of COFL is less than or equal to the computational cost of the FedAvg algorithm, that is,

[0128] cCpu c ≤cCpu a

[0129] prove:

[0130] From equations (17) and (18), we can see

[0131]

[0132] Obviously, in the entire T-round model interaction of federated learning, as long as any edge node does not perform local training, the training cost of COFL is better than that of the FedAvg algorithm.

[0133] Theorem 4. When learning the same model and the edge nodes participating in the training are the same, the time cost of COFL is less than or equal to the time cost of the FedAvg algorithm, that is,

[0134] cTime c ≤cTime a.

[0135] Proof: From equations (19) and (20), we know that

[0136]

[0137] It can be seen intuitively that if the edge node with the maximum training time in any round does not participate in the training, the time cost of COFL is less than or equal to the time cost of the FedAvg algorithm.

[0138] Specifically, in order to verify the effectiveness of the cost-optimized federated learning method, the present invention designs the following algorithm TdCofl based on the COFL model, and will execute it in the experimental part and compare it with the FedAvg algorithm. The process of global model aggregation and local model training is shown in Algorithm 1 and Algorithm 2.

[0139] Aggregation Algorithm for Global Model

[0140]

[0141]

[0142] Among them, edge node 1 is set as the initiator (requester of model training).

[0143] Specifically, in the global model aggregation algorithm (TdCofl-cloud), in any training round, the central coordinator transmits the global model to the edge node, and receives the local model parameters and local model accuracy trained by the edge node, and then calculates the mean and variance of the local model accuracy, and generates a new global model by aggregating the local model parameters within 3 standard deviations, and finally completes T training rounds of federated learning.

[0144] Local training algorithm for edge node k in training round t

[0145]

[0146]

[0147] Among them, E is the number of epochs of local training, and localA represents the verification accuracy of the global model participating in the previous round of training (the initial value is 0).

[0148] Specifically, in the local training algorithm (TdCofl-edge) of training round t, edge node k obtains the global model θ to be trained t, verify the accuracy of the global model. If the verification accuracy is greater than the verification accuracy of the global model participating in the previous round of training, perform local training for E epochs, select the model parameters and model accuracy with the highest local model accuracy and send them to the central coordinator, otherwise set the local model accuracy to 0 and send it to the central coordinator.

[0149] In order to evaluate the communication cost and training cost of the COFL model, the present invention established a Python 3.8 software environment based on Tensorflow2.0, and the hardware environment was 3.60GHz AMD Ryzen 7 3700X 8-core processor CPU, 16.00GB, Win10 64bit, NVIDIA GeForce RTX 2070SUPER.

[0150] The CREDIT_CARD dataset contains 284,807 rows and 31 columns of data. This dataset divides bank accounts into two categories: normal or abnormal based on 30 features, where the data labeled as normal far outnumbers the data labeled as abnormal. In order to realize the characteristics of uneven data distribution in the health crowd intelligence perception scenario, the dataset is randomly divided into 11 parts according to different quantity ratios in the experiment, and randomly distributed to 11 edge nodes. Each node is allocated at least 600 pieces of data to ensure that the validation set has more than 30 samples. The edge node of the model training initiator initializes a 6-layer convolutional neural network model. After the training converges, it submits the training request of federated learning to the central coordinator, which broadcasts it to the 11 edge nodes and starts federated learning training through TdCofl.

[0151] Training Model,A certain edge node acts as the initiator of model training,,initializes a six-layer convolutional neural network model and submits it,along with the training request to the central coordinator.

[0152] In the experiment, the FedAvg algorithm uses the same training model, data set and data partitioning method as TdCofl.

[0153] Since the global model of CREDIT_CARD converges quickly, in order to better compare the communication cost and training cost of TdCofl and FedAvg, this embodiment performs 1000 trainings at T=10, E=5 and T=10, E=2 respectively.

[0154] Next, a detailed analysis is conducted on the experimental results of local model training time, edge node participation in training, communication cost and computing cost, global model accuracy and training loss.

[0155] The duration of local model training in each round and the duration of local model update in federated learning depend on the longest duration of local model training in that round. Figure 3 The average of the longest local model training time in each of the 10 interaction rounds in the experiment is presented. Sum is the total training time of 10 rounds. Obviously, TdCofl significantly saves the training time of the local model compared with FedAvg.

[0156] The status of edge nodes participating in local model training. In each training round of the COFL model, after receiving the global model to be trained transmitted by the central coordinator, the edge node first verifies the global model with the local verification data set. When the verification accuracy is higher than the verification accuracy of the global model that participated in the last training, it participates in the current round of training. Therefore, some edge nodes do not participate in all rounds of local training. Figure 4 The detailed information of the 11 nodes involved in the local model training is displayed. Figure 4 a and b in the equation are two training runs randomly selected from 1000 experiments with E=5. Figure 4 c and d in the figure are 2 training runs randomly selected from 1000 experiments with E=2. Figure 4 The number in the last row summarizes the number of nodes participating in the training in each training round. Since node 1 is the initiator of federated learning, it participates in local training every time. In the figure, y indicates that the edge node participates in the local model training of this round, and n indicates that the edge node does not participate in the local model training of this round.

[0157] Communication cost and computation cost,The comparative analysis of TdCofl and FedAvg proves that the communication cost and computation cost of TdCofl are less than or equal to the communication cost and computation cost of FedAvg.,The two are compared in the experiment, as shown in the following example. Figure 5 shown. Figure 5 In a and c, the horizontal axis is the number of 1000 trials, and the vertical axis is the communication cost of each trial (the size of each uplink model update * the sum of the number of nodes participating in training in 10 rounds). Obviously, the communication cost of TdCofl is much better than that of FedAvg. Since the global model communication cost from the central coordinator to the edge node is the same for TdCofl and FedAvg, only the traffic from the edge node uploading the model update to the central coordinator is compared here. The horizontal axis of b and d is 10 training rounds, and the vertical axis is the average number of nodes participating in local model training in each round of 1000 experiments. The edge nodes that did not participate in local model training saved computational costs. It can be seen from the figure that the computational cost of TdCofl is much better than that of FedAvg, and because the global model converges slower when E=2 than when E=5, the number of nodes participating in training increases and the communication cost increases.

[0158] The training accuracy and training loss of the global model, Figure 6The changes in the global model accuracy of TdCofl and FedAvg in 1000 trials of E=5 and E=2 are presented, where the horizontal axis is the training round and the vertical axis is the model accuracy. Obviously, TdCofl and FedAvg have similar model accuracy and training loss, and TdCofl tends to be more stable.

[0159] Based on the above analysis of the experimental results, it can be seen that the experimental results further verify the theoretical analysis of the COFL model in this embodiment. Compared with FedAvg, the cost-optimized federated learning method can reduce the communication cost of federated learning while ensuring training accuracy and reducing training costs.

[0160] In summary, starting from the unique attributes of health crowd intelligence perception, the present invention introduces the service paradigm of cloud-edge architecture for the first time, deploys federated learning to extract group intelligence in view of the strong privacy characteristics in the health field; innovatively improves the federated learning process to reduce the model training cost and parameter transmission cost in the federated learning process, and effectively improves the aggregation efficiency of the global model. The performance of the method proposed in the present invention is compared and analyzed with the FedAvg algorithm, which proves that the method of the present invention reduces the training cost while reducing the communication cost. Through a large number of experiments based on open data sets, it is shown that the communication cost, local model training time and computational cost of the method of the present invention are superior to the FedAvg algorithm, which verifies the effectiveness of the proposed method.

[0161] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A cost-optimized federated learning method in a health crowd intelligence perception system, characterized in that: include: Establish a cost-optimized federated learning model based on the goals of federated learning; Before each local training, the edge node verifies the received global model with the verification data set in the real-time data. If its verification accuracy is higher than the verification accuracy of the previous round, it participates in this training, otherwise it refuses to participate in this training; When edge nodes perform local training in each communication round, they conduct multiple epochs of training and select the model parameters with the highest verification accuracy and upload them to the central coordinator for model aggregation; After receiving the local model parameters from the edge nodes, the central coordinator uses the truth value discovery method to evaluate the quality of the local model parameters and selects the local model parameters that meet the quality evaluation requirements to participate in the aggregation of the global model; The cost-optimized federated learning method in the health crowd intelligence perception system adopts a health crowd intelligence perception system, which includes: a central coordinator, an edge node and an end user; the central coordinator is a health department or a hospital consortium; the edge node is a hospital or a community medical center; the end user is an intelligent terminal carried by the patient, which is used to collect patient data and upload the patient data to the edge node; The central coordinator and edge nodes jointly deploy federated learning, which implements intelligent extraction with privacy protection. Edge nodes and end users collaborate in data collection and management.

2. According to claim 1, a cost-optimized federated learning method in a health crowd intelligence perception system is characterized in that: The service process of the health crowd intelligence perception system is as follows: Any edge node acts as the initiator of model training, selects a training model as needed, initializes the model using local data, and submits the model training requirements and the initialized model to the central coordinator; The central coordinator distributes the global model to be trained to the edge nodes participating in the training, obtains the local model parameters of the edge nodes after training, performs quality assessment on the model parameters, selects the local model parameters participating in the global model aggregation based on the quality assessment results, and generates a new global model; The edge node stores the data uploaded by the smart terminal, receives and verifies the global model issued by the central coordinator, and as the data holder, trains the global model with local data to generate a new local model and uploads the model parameters to the central coordinator; Patients perceive and generate individual in-depth data through the smart terminals they carry, upload the data to edge nodes, and receive personalized diagnosis and treatment services.

3. The cost-optimized federated learning method in a health crowd intelligence perception system according to claim 1, characterized in that: The cost-optimized federated learning model established according to the goal of federated learning includes: According to the goal of federated learning, in f:R d →R is the global loss function, f k :R d →R is the local loss function of edge node k, θ∈R d is the optimizable model parameter, θ * is the optimized model parameter, p k is the ratio of the amount of training data of edge node k to the total amount of training data participating in the global model aggregation, l(θ,x) is the loss function of data sample x when the model is θ, {1,2,...,K} is the set of edge nodes trained by the federated learning model, k is any edge node, and D k and |D k | are the training data set and training data volume of the kth edge node respectively; Building a cost-optimized federated learning model: Where d is the sum of the training data of the edge nodes where the local model parameters participating in the global model aggregation are located in the current training round, and d K is the sum of the training data of the edge nodes participating in the training, p k and are the ratio of the amount of training data of the kth edge node to the total amount of training data participating in the global model aggregation and to the total amount of overall training data, θ t represents the global model of the tth round, represents the local model of the kth edge node in the tth round, a k is the model accuracy of the kth edge node, μ and σ are the mean and standard deviation of the accuracy of all local models participating in the training, respectively.

Images