Federal crowd sensing method for online learning

Abstract

The invention relates to a federated crowd sensing method for online learning, and belongs to the technical field of Internet of Things. The implementation method comprises the following steps: 1, constructing a perception interaction model of a server and a participant; modeling interaction modes and effectiveness of the participants and the servers and optimizing the interaction modes of the participants and the servers; furthermore, the perception tasks are optimally distributed in a way of reward budget and payment for the interaction mode; the diversity of sensing task data acquisition is improved; 2, constructing a federal interaction model of the server and the participant; in a random and antagonistic environment, bidding is performed on modeling participants, the selection probability of the participants is optimized, the selection probability estimator is updated through bidding of the participants and contribution to a global model, and then the federated learning model is optimized; compared with the prior art, the method solves the problem of diversity and privacy of data in existing Internet of Things data analysis in random and antagonistic environments, and further realizes efficient Internet of Things data analysis.

Description

Technical Field

[0001] The invention relates to a federated group intelligence perception method for online learning, and belongs to the technical field of Internet of Things. Background Art

[0002] In recent years, the development of IoT and AI technologies has spawned applications such as autonomous driving, smart health, and smart home in various industries. Among these application services, data analysis is one of the most critical links. Data analysis relies on data element resources, especially various types of data resources are needed to support complex intelligent application services. Therefore, the comprehensive and multi-angle perception and collection of various types of data resources are crucial to the design and development of downstream applications.

[0003] With the widespread popularity of mobile devices, these mobile devices usually have processors with good performance and large storage space. They also have built-in multiple types of sensors, such as GPS, cameras, air pressure sensors, accelerometers, etc., which can sense the surrounding environment at all times and generate a large amount of sensor data. However, these sensor data and the computing resources of mobile devices are often idle and not fully utilized. Therefore, crowd sensing technology came into being. Crowd sensing technology can be used to obtain data purposefully and increase the perception efficiency of useful information. At the same time, data resources are private and cannot be uploaded to the service platform, so user participants are required to perform tasks locally. As a distributed learning technology with privacy protection function, federated learning technology allows global prediction models to be trained without leaving the domain, thereby solving the problem of privacy protection in crowd sensing technology. However, the combination of crowd sensing and federated learning alone can provide diversity and security and privacy protection for data, but it cannot be applied to random and adversarial environments. For example, the user participant training process uses stochastic gradient descent to make the model training random. In addition, user participants are adversarial in order to compete for participation, which brings challenges to selecting high-quality user participants.

[0004] Therefore, in a random and adversarial environment, how to ensure data diversity and privacy in existing IoT data analysis becomes an urgent problem to be solved. Summary of the invention

[0005] The purpose of the present invention is to solve the technical problems of data diversity and privacy in existing IoT data analysis in random and adversarial environments, and to propose an online learning federated crowd intelligence perception method.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] An online learning federated group intelligence perception method of the present invention is applied to a task node including a server member node and multiple participant member nodes, wherein the server member node is used for requesting and assigning perception tasks and federated learning tasks; the participant member node is used for performing task perception and feeding back local model parameters to the server member node;

[0008] Step 1: Construct the perception interaction model of the server member node and the participant member node; model the interaction mode and utility of the participant member node and the server member node respectively and optimize the interaction mode of both parties; further optimize the allocation of perception tasks by using reward budget and payment for the interaction mode; thereby improving the diversity of perception task data collection;

[0009] Step 1.1: Model the participant member nodes and optimize the perception quantity; model the server member nodes and optimize the reward budget and task announcement;

[0010] Step 1.1.1: Construct the utility model of the participant member nodes for crowd sensing in the manner shown in formula (1);

[0011]

[0012] Among them, u i represents the utility value of participant member node i, n i represents the perceived quantity, R represents the reward budget of the server member node, and k i is the perceived cost of participant member node i, is the set of participant member nodes;

[0013] Step 1.1.2: Optimize the perceived quantity of the participant using the method shown in formula (2); and then obtain the perceived quantity of the participant node;

[0014]

[0015] Among them, the collection represents all other participant member nodes except participant member node i;

[0016] Step 1.1.3: Use the perceived number of participant member nodes to model the utility of the server member nodes in the manner shown in formula (3);

[0017]

[0018] Among them, λ represents the benefit gain of the perceived quantity of all participant member nodes to the service provider member nodes, and the log(·) function is used to characterize the characteristics of diminishing marginal benefits;

[0019] Step 1.1.4: Use Newton's method to optimize the reward budget R of the server member node;

[0020] Step 1.1.5: The server member node announces the sensing task and reward budget;

[0021] Step 1.2: The participant member node uses the sensing quantity to sense and stores the sensing data in the local database of the participant member node; the server member node receives the sensing task report of the participant member node and pays the participant member node according to the reward budget;

[0022] Step 2: Construct a federated interaction model of the server member nodes and the participant member nodes; in a random and adversarial environment, model the bids of the participant member nodes and optimize the selection probability of the participant member nodes, update the selection probability estimator through the bids of the participant member nodes and their contribution to the global model, and then optimize the federated learning model;

[0023] Step 2.1: In a random and adversarial environment, construct the bids of the participant's member nodes, use the bids to estimate the cumulative loss reward vector, and then construct a relationship model between selection probability and bids; generate a list of participant member nodes to be executed by optimizing the selection probability;

[0024] Step 2.1.1: During a round of federated learning, the participating member nodes bid r in a random and adversarial manner. i,t ;

[0025] Step 2.1.2: Use the method shown in formula (4) to obtain the relationship model between selection probability and bid, and then obtain the probability of the participant member node being selected;

[0026] The loss estimate for a single round is given by vector representation, ψ(x) represents the regularizer, η t represents the time-decayed learning rate;

[0027] The specific ψ(x) is set as shown in formula (5); η t The setting is performed in the manner shown in formula (6);

[0028]

[0029] Step 2.1.3: Using the probability x of being selected t Select the participant member nodes; and then generate a list of the participant member nodes to be executed;

[0030] Step 2.2: Perform local training on the member nodes of the task participants in the federated learning to obtain the accuracy contribution of the member nodes of the task participants to the global model. i,t , pay the reward r according to the bid of the member node of the task to be executed i,t At the same time, the bids of the participant member nodes and the accuracy contribution of the global model are aggregated using formula (7) to obtain the aggregated observation l t,i ;

[0031]

[0032] The hyperparameter β is a regulator of the bid and global model contribution, making l t,i The aggregate observation l of the participant member nodes in the list of participant member nodes of the task to be executed falls in the interval [-1,1]. t,i =0;

[0033] Step 2.3: Obtain the cumulative loss statistics for the next round of federated learning using the method shown in formula (8);

[0034]

[0035] in, It is expressed as a single-round loss estimator, whose i-th dimension It is constructed through the loss function shown in formula (9);

[0036]

[0037] Among them, t (i) indicates whether participant member node i is selected in round t;

[0038] Step 2.4: Execute steps 2.1 to 2.3 in a loop iteration manner to form an optimized global model;

[0039] Beneficial effects:

[0040] Compared with the prior art, the present invention has the following effects:

[0041] 1. Compared with the existing technology, the method of the present invention combines the advantages of crowd sensing and federated learning, while overcoming the privacy protection problem of traditional crowd sensing computing and the problem of effective data source of traditional federated learning, increasing the diversity of training data, improving the generalization performance of the global model in multiple data types, and protecting the privacy of the data.

[0042] 2. In addition, the present invention designs a federated crowd intelligence perception method based on online learning, which solves the problem that it is difficult to identify high-quality participant member nodes in random and adversarial environments. It minimizes the cumulative regret value of the task execution process through online learning technology, eliminates the uncertainty caused by random and adversarial environments, and realizes efficient IoT data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the federated group intelligence perception system according to an embodiment of the present invention;

[0044] Figure 2 A graph showing changes in the number of user participants in the perception phase of an embodiment of the present invention;

[0045] Figure 3 , Figure 4 A perception platform benefit diagram of the perception stage of an embodiment of the present invention;

[0046] Figure 5 , Figure 6 A diagram showing changes in the optimal reward budget in the perception phase of an embodiment of the present invention;

[0047] Figure 7 A graph showing changes in the accumulated regret value during the federated learning phase of an embodiment of the present invention;

[0048] Figure 8 This is a graph showing changes in the cumulative contribution during the federated learning phase of an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to better illustrate the purpose and advantages of the present invention, the invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that the implementation of the present invention is not limited to the following embodiments, and any form of modification or change made to the present invention will fall within the protection scope of the present invention.

[0050] Example

[0051] like Figure 1 As shown, the federated crowd intelligence perception method of online learning in this embodiment has the following specific implementation steps:

[0052] Step 1: Construct the perception interaction model of the server member node and the participant member node; model the interaction mode and utility of the participant member node and the server member node respectively and optimize the interaction mode of both parties; further optimize the allocation of perception tasks by using reward budget and payment for the interaction mode; thereby improving the diversity of perception task data collection;

[0053] Step 1.1: Model the participant member nodes and optimize the perception quantity; model the server member nodes and optimize the reward budget and task announcement;

[0054] Step 1.1.1: Construct the utility model of the participant member nodes for crowd sensing in the manner shown in formula (1);

[0055]

[0056] Among them, u i represents the utility value of participant member node i, n i represents the perceived quantity, R represents the reward budget of the server member node, and k i is the perceived cost of participant member node i, is the set of participant member nodes;

[0057] Step 1.1.2: Optimize the perceived quantity of the participant using the method shown in formula (2); and then obtain the perceived quantity of the participant node;

[0058]

[0059] Among them, the collection represents all other participant member nodes except participant member node i;

[0060] Step 1.1.3: Use the perceived number of participant member nodes to model the utility of the server member nodes in the manner shown in formula (3);

[0061]

[0062] Among them, λ represents the benefit gain of the perceived quantity of all participant member nodes to the service provider member nodes, and the log(·) function is used to characterize the characteristics of diminishing marginal benefits;

[0063] Step 1.1.4: Use Newton's method to optimize the reward budget R of the server member node;

[0064] Step 1.1.5: The server member node announces the sensing task and reward budget;

[0065] Step 1.2: The participant member node uses the sensing quantity to sense and stores the sensing data in the local database of the participant member node; the server member node receives the sensing task report of the participant member node and pays the participant member node according to the reward budget;

[0066] In this embodiment, it is assumed that the perceived cost of each participant member node is uniformly distributed in [1,k max ], where k max The hyperparameter λ in equation (3) is set to 20. The total number of user participants, i.e., the set The size of the data was increased from 100 to 1000. In order to eliminate the influence of data randomness, the average value of the data from 5 experiments was taken. Figure 2 The figure shows the change in the number of participating users in the perception phase of the method of the present invention. It can be observed that, when the total number of participant member nodes is fixed, the number of participant member nodes will decrease as the unit cost of the participant member nodes increases.

[0067] Figure 3 and 4 Shows the number of participant member nodes and the maximum perceived cost k of the participant member nodes max Impact on perceived platform benefits. Figure 3 In the example, k is set max = 5 and conduct experiments. It can be observed that the perceived platform benefits do show an increase with The return is decreasing with the increase of Figure 4 In the It can be observed that as the unit cost of the participant member node increases, the perceived platform benefit decreases.

[0068] Figure 5 and Figure 6 The number of participant member nodes is shown separately and the maximum perceived cost k of the participant member nodes max Impact on the optimal reward budget of the server member nodes. Figure 5 It can be observed that the optimal reward budget R * The value of increases with the increase of The increase gradually stabilizes. Figure 6 It can be observed that as the maximum perceived cost k of the participant member node max The increase in the optimal reward budget R of the server member node * reduce.

[0069] Step 2: Construct a federated interaction model of the server member nodes and the participant member nodes; in a random and adversarial environment, model the bids of the participant member nodes and optimize the selection probability of the participant member nodes, update the selection probability estimator through the bids of the participant member nodes and their contribution to the global model, and then optimize the federated learning model;

[0070] Step 2.1: In a random and adversarial environment, construct the bids of the participant's member nodes, use the bids to estimate the cumulative loss reward vector, and then construct a relationship model between selection probability and bids; generate a list of participant member nodes to be executed by optimizing the selection probability;

[0071] Step 2.1.1: During a round of federated learning, the participating member nodes bid r in a random and adversarial manner. i,t ;

[0072] Step 2.1.2: Use the method shown in formula (4) to obtain the relationship model between selection probability and bid, and then obtain the probability of the participant member node being selected;

[0073]

[0074] The loss estimate for a single round is given by vector representation, ψ(x) represents the regularizer, η t represents the time-decayed learning rate;

[0075] The specific ψ(x) is set as shown in formula (5); η t The setting is performed in the manner shown in formula (6);

[0076]

[0077] Step 2.1.3: Using the probability x of being selected t Select the participant member nodes; and then generate a list of the participant member nodes to be executed;

[0078] Step 2.2: Perform local training on the member nodes of the task participants in the federated learning to obtain the accuracy contribution of the member nodes of the task participants to the global model. i,t , pay the reward r according to the bid of the member node of the task to be executed i,t At the same time, the bids of the participant member nodes and the accuracy contribution of the global model are aggregated using formula (7) to obtain the aggregated observation l t,i ;

[0079]

[0080] The hyperparameter β is a regulator of the bid and global model contribution, making l t,i The aggregate observation l of the participant member nodes in the list of participant member nodes of the task to be executed falls in the interval [-1,1]. t,i =0;

[0081] Step 2.3: Obtain the cumulative loss statistics for the next round of federated learning using the method shown in formula (8);

[0082]

[0083] in, It is expressed as a single-round loss estimator, whose i-th dimension It is constructed through the loss function shown in formula (9);

[0084]

[0085] Among them, t (i) indicates whether participant member node i is selected in round t;

[0086] Step 2.4: Execute steps 2.1 to 2.3 in a loop iteration manner to form an optimized global model;

[0087] In the embodiment, the cumulative pseudo regret is used And cumulative contribution To calculate the online learning performance of this step, assume that the bid of participant member node i at the beginning of the first round is r 1,i is a random number between (0,1). In each subsequent round, if it is not selected by the platform, it will be reduced by 1 / 4, and if it is selected, it will be increased by 1 / 3. The maximum value is no more than 2 times the initial random value, and the minimum value is no less than 1 / 2 of the initial random value. The contribution of participant member node i to the global model in each round is random number a t,i , is a set of random numbers distributed on (0,1) and generated at the beginning of the simulation. Set the values ​​of γ and β to 1. Number of users The number of users selected in each group is d = 20. CombUCB algorithm, CombEXP3 algorithm and ThompsonSampling algorithm are used as the control group.

[0088] Figure 7 The cumulative regret value of this method is shown in Figure 7 It can be seen that as the number of rounds increases, the cumulative regret value generated by this method will decrease. Compared with the other two algorithms, this method can make the cumulative regret converge quickly before the budget is used up. When too many rounds of training cannot be performed due to budget constraints, this method performs better. Figure 8 The cumulative contribution of this method is shown in Figure 2. Figure 8 It can be seen that under the same experimental conditions, the performance of this method is almost equal to that of other algorithms. Therefore, under the same conditions, this method has better overall performance.

[0089] The specific description above further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An online learning federated crowd intelligence perception method, characterized by: The following steps are included: Step 1: Construct the perception interaction model of the server member node and the participant member node; model the interaction mode and utility of the participant member node and the server member node respectively and optimize the interaction mode of both parties; further optimize the allocation of perception tasks by using reward budget and payment for the interaction mode; thereby improving the diversity of perception task data collection; Step 1.1: Model the participant member nodes and optimize the perception quantity; model the server member nodes and optimize the reward budget and task announcement; Step 1.2: The participant member node uses the sensing quantity to sense and stores the sensing data in the local database of the participant member node; the server member node receives the sensing task report of the participant member node and pays the participant member node according to the reward budget; Step 2: Construct a federated interaction model of the server member nodes and the participant member nodes; in a random and adversarial environment, model the bids of the participant member nodes and optimize the selection probability of the participant member nodes, update the selection probability estimator through the bids of the participant member nodes and their contribution to the global model, and then optimize the federated learning model; Step 2.1: In a random and adversarial environment, construct the bids of the participant's member nodes, use the bids to estimate the cumulative loss reward vector, and then construct a relationship model between selection probability and bids; generate a list of participant member nodes to be executed by optimizing the selection probability; Step 2.2: Perform local training on the member nodes of the task participants in the federated learning to obtain the accuracy contribution of the member nodes of the task participants to the global model. i,t , pay the reward r according to the bid of the member node of the task to be executed i,t At the same time, the bids of the participant member nodes and the accuracy contribution of the global model are aggregated using formula (7) to obtain the aggregated observation l t,i ; The hyperparameter β is a regulator of the bid and global model contribution, making l t,i The aggregate observation l of the participant member nodes in the list of participant member nodes of the task to be executed falls in the interval [-1,1]. t,i =0; Step 2.3: Obtain the cumulative loss statistics for the next round of federated learning using the method shown in formula (8); in, It is expressed as a single-round loss estimator, whose i-th dimension It is constructed through the loss function shown in formula (9); Among them, t (i) indicates whether participant member node i is selected in round t; Step 2.4: Execute steps 2.1 to 2.3 in an iterative manner to form an optimized global model.

2. The method for federated crowd intelligence perception of online learning according to claim 1, characterized in that: Step 1.1 is implemented as follows: Step 1.1.1: Construct the utility model of the participant member nodes for crowd sensing in the manner shown in formula (1); Among them, u i represents the utility value of participant member node i, n i represents the perceived quantity, R represents the reward budget of the server member node, and k i is the perceived cost of participant member node i, is the set of participant member nodes; Step 1.1.2: Optimize the perceived quantity of the participant using the method shown in formula (2); and then obtain the perceived quantity of the participant node; Among them, the collection represents all other participant member nodes except participant member node i; Step 1.1.3: Use the perceived number of participant member nodes to model the utility of the server member nodes in the manner shown in formula (3); Where λ represents the benefit gain of the perceived quantity of all participant member nodes to the service provider member nodes, the log(·) function is used to characterize the characteristics of diminishing marginal benefits, and the other symbols have the same meaning as formula (1); Step 1.1.4: Use Newton's method to optimize the reward budget R of the server member node; Step 1.1.5: The server member node publishes the sensing task and reward budget.

3. The method for federated crowd intelligence perception of online learning according to claim 1, characterized in that: Step 2.1 is implemented as follows: Step 2.1.1: During a round of federated learning, the participating member nodes bid r in a random and adversarial manner. i,t ; Step 2.1.2: Use the method shown in formula (4) to obtain the relationship model between selection probability and bid, and then obtain the probability of the participant member node being selected; Among them, t represents the current round, Represents the cumulative reward loss vector, which is initialized to The loss estimate for a single round is given by vector representation, ψ(x) represents the regularizer, η t represents the time-decayed learning rate; The specific ψ(x) is set as shown in formula (5); η t The setting is performed in the manner shown in formula (6); Step 2.1.3: Using the probability x of being selected t Select the participant member nodes; and then generate a list of participant member nodes for tasks to be executed.

Images