Efficient federated learning method based on cloud-edge in smart home scenarios

Abstract

The present invention provides an efficient cloud-edge-device-based federated learning method in a smart home scenario. The method comprises: a cloud server acquires a learning task and initializes an initial model, and sends it to a communication base station participating in the training; the communication base station sends the learning task and initial model to a temporary server and home terminal participating in the training; the home terminal performs local training based on local data and the initial model or a first intermediate model, obtains a new local model, and transmits it to the temporary server; the temporary server performs weighted aggregation on all local models and its own local model to obtain a first intermediate model and returns it to each home terminal; the first two steps are repeated until an end condition is met, and the first intermediate model is transmitted to the communication base station; the communication base station performs weighted aggregation on all first intermediate models to obtain a second intermediate model and returns it to the cloud server; the cloud server aggregates all second intermediate models to obtain a final global model.

Description

Technical Field

[0001] The present invention relates to the field of smart home and privacy security technologies, and in particular to an efficient cloud-edge-based federated learning method in a smart home scenario. Background Art

[0002] In recent years, the development of the Internet of Things (IoT) has been geared towards providing greater comfort and convenience for people, and home products are no exception. For example, home cameras can monitor homes for fire hazards, robot vacuums can help with hygiene, and smart meters can monitor the power usage of household appliances and optimize energy-saving modes. As everyday objects, home products are becoming increasingly intelligent with the development of the internet. As these intelligent home products provide higher-quality services, manufacturers need to use these digital products to collect daily life data and optimize service models. However, this daily life data is often relatively private to users (such as voice interaction data, electricity usage data, and home profile data). Therefore, while training models for smart homes, smart homes also face the risk of privacy data leakage.

[0003] Federated learning solves many data security issues by ensuring data security by keeping the original dataset local. This concept is being applied in the smart home sector, allowing models to be trained locally on smart home devices and then transmitting the updated data to a central server on the home unit's local area network for further model aggregation. By keeping the data local, data security is significantly improved.

[0004] However, due to the clustered nature of residential communities in my country, many challenges remain. First, with the continuous development of smart home products, device manufacturers are using the cloud to communicate directly with home devices in densely populated areas. This not only increases the pressure on the cloud for model training, but also increases the data communication load. Second, smart home devices have varying computing power and the characteristics of the data they collect are also inconsistent, which leads to some challenges in model aggregation. Finally, the idleness of home terminals is relatively flexible, while existing federated learning algorithms are at a disadvantage in home scenarios with uneven computing power distribution and large data volumes.

[0005] Chinese patent document CN 113946127 A discloses a smart home system based on edge computing technology. In this solution, the algorithm trained in the smart home scenario is an average aggregation method of a single model. It does not interact with the cloud server and requires a separate device to be designated as the central server. In addition, this method is costly and time-consuming during the model training process. Therefore, it is not helpful for improving computing efficiency in dense terminal devices in community environments. Summary of the Invention

[0006] In order to improve the efficiency and accuracy of model training in smart home scenarios, the present invention provides an efficient cloud-edge-end federated learning method in smart home scenarios.

[0007] The present invention provides an efficient cloud-edge-device federated learning method in a smart home scenario, comprising:

[0008] Step 1: The cloud server obtains the learning task and initializes the initial model;

[0009] Step 2: The cloud server sends the learning task and initial model to the communication base stations participating in the training;

[0010] Step 3: The communication base station sends the acquired learning task and initial model to the temporary server and home terminals in the home unit participating in the training within its communication range; the temporary server is elected by all home terminals in a home unit;

[0011] Step 4: The home terminal performs local model training and updates based on local data and the received initial model or the first intermediate model of the previous round to obtain a new local model, and transmits the local model to the temporary server;

[0012] Step 5: The temporary server performs weighted aggregation on all received local models and its own local model to obtain a first intermediate model, and returns the first intermediate model to each home terminal;

[0013] Step 6: Repeat steps 4 to 5 until the end condition is met, and the temporary server transmits the first intermediate model to the communication base station;

[0014] Step 7: The communication base station performs weighted aggregation on all received first intermediate models to obtain a second intermediate model, and returns the second intermediate model to the cloud server;

[0015] Step 8: The cloud server aggregates all received second intermediate models to obtain the final global model.

[0016] Furthermore, the election of the temporary server includes: electing a temporary server according to the computing capabilities and idle conditions of all home terminals in the home unit.

[0017] Furthermore, a temporary server is selected based on the computing power and idleness of all home terminals in the home unit, specifically including:

[0018] Calculate the sum of computing power and idle time of each home terminal;

[0019] The home terminal with the largest sum of computing power and idle time is selected as the temporary server of the home unit.

[0020] Furthermore, it also includes:

[0021] The temporary server collects local data from home terminals without computing power in the home unit, and combines its own local data with the received initial model or the first intermediate model of the previous round to perform local model training and update to obtain its own new local model.

[0022] Furthermore, before step 2, the method further includes:

[0023] For the communication base stations that need to participate in the training, the communication base stations are registered with the cloud server, so that the cloud server can collect statistics on all the communication base stations and form a base station index;

[0024] Correspondingly, step 2 is specifically as follows: the cloud server sends the learning task and the initial model to the corresponding communication base station according to the base station index.

[0025] Furthermore, before step 3, the method further includes:

[0026] For the home unit that needs to participate in the training, the temporary server and / or home terminal of the home unit is registered with its corresponding communication base station, so that the communication base station can collect statistics on all temporary servers and / or home terminals and form a temporary server index;

[0027] Correspondingly, step 3 is specifically as follows: the communication base station sends the acquired learning tasks and initial models to the corresponding temporary server and home terminal according to the temporary server index.

[0028] Furthermore, step 4 specifically includes:

[0029] Home terminal i calculates the gradient descent of the tth iteration process based on local data and the received initial model or the first intermediate model of the previous round Where f(ω)=l(ω,D), f(ω) represents the prediction loss function, ω represents the parameters of the initial model received by home terminal i or the first intermediate model of the previous round, and D represents the local data of home terminal i. represents the gradient operator;

[0030] Home terminal i descends according to the gradient Update the local model of the tth iteration process Among them, η i For the learning efficiency of home terminal i, Represents the parameters of the local model of home terminal i before the local training of the t-th iteration process.

[0031] Furthermore, step 5 specifically includes:

[0032] The temporary server performs weighted aggregation on each local model according to formula (1) to obtain the parameters of the first intermediate model:

[0033] ω j+1 =ω j +ω i (1)

[0034] Among them, ω j+1 represents the parameters of the first intermediate model obtained by the temporary server j in the current round of aggregation, ω j represents the parameters of the first intermediate model obtained after the last round of aggregation of temporary server j, ω i Represents the current local model of home terminal i.

[0035] Furthermore, step 7 specifically includes:

[0036] For each communication base station, all parameters of the first intermediate model are added together to obtain ω s According to the total number of temporary servers participating in the training in the communication base station, s Perform iterative averaging, and use the iterative averaging results as the parameters of the second intermediate model.

[0037] Furthermore, the local data of the home terminal includes at least one of network throughput, power usage and camera monitoring data.

[0038] Beneficial effects of the present invention:

[0039] (1) The number of smart home terminals in a household is exploding. If the training data of each home terminal in a unit household is transmitted to the cloud server, the bandwidth requirement is relatively high. The efficient federated learning method designed in the cloud-edge architecture of the present invention can reduce the communication load on the cloud server compared to the direct communication between the terminal device and the cloud server.

[0040] (2) When the model is aggregated, the aggregation process requires corresponding computing resources. A single terminal directly transmits the model results to the cloud server to participate in the aggregation. If the number of terminals participating in the aggregation is large, the cloud server needs to have higher computing resources. Through the cloud-edge model aggregation method designed by the present invention, the aggregation pressure of the model is dispersed to the temporary server and the communication base station, reducing the aggregation pressure of the cloud server, improving the overall training efficiency, and enhancing the efficiency of learning of smart home terminals. At the same time, the accuracy of training is also guaranteed during the repeated iteration process within the home unit;

[0041] (3) The information perceived by a single household terminal device is limited, but by training a model and aggregating the perceived information of multiple households, more accurate and comprehensive data information can be obtained; therefore, the present invention can make full use of the data perceived by the smart home terminal devices;

[0042] (4) The temporary server of the family is selected based on the computing power and idleness of the terminal devices in the family unit, which can reduce the cost of setting up a separate server in the family, make better use of idle resources, and save some costs;

[0043] (5) A joint training model is established between smart home terminals through the communication base station and the cloud server. The smart home terminals do not need to upload local data to the communication base station and the cloud server. Instead, they update the minimum model trained locally by the smart home terminals and upload it to the communication base station and the cloud server for model aggregation. When the data does not leave the local smart home terminal, the privacy of the family's own data is protected. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the process of an efficient cloud-edge-device federated learning method in a smart home scenario provided by an embodiment of the present invention;

[0045] Figure 2 A diagram of information interaction between a home terminal and a temporary server, between a temporary server and a communication base station, and between a communication base station and a cloud server provided in an embodiment of the present invention;

[0046] Figure 3 This is a diagram showing privacy changes of a home terminal under different methods provided by an embodiment of the present invention;

[0047] Figure 4 A graph showing changes in learning efficiency of the method provided by the embodiment of the present invention and two existing traditional methods;

[0048] Figure 5 A time accuracy change diagram of the method provided in the embodiment of the present invention and two existing traditional methods. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] Example 1

[0051] like Figure 1 As shown, an embodiment of the present invention provides an efficient federated learning method based on cloud-edge in a smart home scenario, including:

[0052] S101: The cloud server obtains the learning task and initializes the initial model;

[0053] S102: The cloud server sends the learning task and the initial model to the communication base stations participating in the training;

[0054] S103: The communication base station sends the acquired learning task and initial model to the temporary server and home terminals in the home unit participating in the training within its communication range; the temporary server is elected by all home terminals in a home unit;

[0055] S104: The home terminal performs local model training and updating based on the local data and the received initial model or the first intermediate model of the previous round to obtain a new local model, and transmits the local model to the temporary server;

[0056] S105: The temporary server performs weighted aggregation on all received local models and its own local model to obtain a first intermediate model, and returns the first intermediate model to each home terminal;

[0057] S106: Repeat steps S104 to S105 until the termination condition is met, and the temporary server transmits the first intermediate model to the communication base station;

[0058] S107: The communication base station performs weighted aggregation on all received first intermediate models to obtain a second intermediate model, and returns the second intermediate model to the cloud server;

[0059] S108: The cloud server aggregates all received second intermediate models to obtain a final global model.

[0060] The embodiment of the present invention is based on the idea of ​​cloud-edge federated architecture, and uses communication base stations (BS) and temporary servers (LTS) in home units to perform hierarchical aggregation and step-by-step aggregation to disperse the data aggregation pressure to the home unit and the communication base station, and make full use of idle resources in the home unit. Therefore, the present invention can reduce communication and computing overhead in intensive situations, optimize the consumption in data communication and computing, and thus improve the overall learning efficiency; and the global model obtained through three aggregations also ensures the accuracy of the model.

[0061] Example 2

[0062] Based on the above embodiment 1, this embodiment of the present invention provides another efficient cloud-edge-device-based federated learning method in a smart home scenario, including the following steps:

[0063] S201: The cloud server obtains the learning task and initializes the initial model;

[0064] The learning task may be a learning task assigned by a manufacturer corresponding to the smart home terminal.

[0065] S202: For the communication base stations that need to participate in the training, register the communication base stations with the cloud server, so that the cloud server can collect statistics on all the communication base stations and form a base station index B. i , i=1,2,...,M;B i Represents the index value of communication base station i;

[0066] S203: The cloud server sends the learning task and the initial model to the corresponding communication base station according to the base station index;

[0067] S204: Selecting a temporary server in the home unit based on the computing power and idle status of all home terminals in the home unit; specifically including: calculating the sum of the computing power and idle time of each home terminal; and selecting the home terminal with the largest sum of computing power and idle time as the temporary server of the home unit.

[0068] S205: For the family unit that needs to participate in the training, the temporary server and / or home terminal of the family unit is registered with its corresponding communication base station, so that the communication base station can count all the temporary servers and / or home terminals and form a temporary server index C j , j=1,2,...,N;C j Indicates the index value of temporary server j.

[0069] In this embodiment, on the one hand, the temporary server can determine whether to participate in the training process of the learning task based on the idle status of other home terminals in its own family unit and / or the rewards brought by the learning task; on the other hand, each home terminal in the family unit can also actively choose to participate in the training process of the learning task.

[0070] S206: The communication base station sends the acquired learning task and initial model to the corresponding temporary server and home terminal according to the temporary server index;

[0071] S207: The home terminal performs local model training and updating based on the local data and the received initial model or the first intermediate model of the previous round to obtain a new local model, and transmits the local model to the temporary server;

[0072] Taking the local model training process of a home terminal i as an example, the home terminal i calculates the gradient descent of the tth iteration process based on the local data and the received initial model or the first intermediate model of the previous round Where f(ω)=l(ω,D), f(ω) represents the prediction loss function, ω represents the parameters of the initial model received by home terminal i or the first intermediate model of the previous round, and D represents the local data of home terminal i. represents the gradient operator;

[0073] Home terminal i descends according to the gradient Update the local model of the tth iteration process Among them, η i For the learning efficiency of home terminal i, Represents the parameters of the local model of home terminal i before the local training of the t-th iteration process.

[0074] It is understandable that during the first round of training, the model involved in the training is the initial model. The local data involved in the training includes network throughput, power usage, camera monitoring, and other data.

[0075] S208: The temporary server collects local data from home terminals without computing capabilities in the home unit, and combines the local data with its own local data and the initial model or the first intermediate model of the previous round to perform local model training and update, thereby obtaining its own new local model;

[0076] The local model training process of the temporary server is the same as the local model training of the home terminal in the previous step, and will not be repeated here.

[0077] Understandably, home devices without computing power also contain some important data. Therefore, to obtain more household life data and improve model accuracy, a temporary server is responsible for collecting data from these devices locally. Furthermore, the temporary server can convert the operating status of sensor devices without computing power into usable data; this data is then combined with the temporary server's own local data for training. For example, a home smart TV acts as a temporary server. Part of the local training data comes from the TV's own historical usage data, while part comes from smart lighting data, camera data, and other data within the home unit's local area network. This data is aggregated into the TV's local data for training, resulting in a more accurate model.

[0078] Depending on the needs of different tasks, data with different features are used for model training, such as detecting whether a fire has occurred through home cameras and other image data.

[0079] S209: The temporary server performs weighted aggregation on all received local models and its own local model to obtain a first intermediate model, and returns the first intermediate model to each home terminal;

[0080] In this embodiment, in order to improve the aggregation speed of the model, the temporary server aggregates the received local models into the first intermediate model in real time, rather than waiting until all local models are collected before aggregating; that is, the temporary server weights and aggregates each local model according to formula (1) to obtain the parameters of the first intermediate model:

[0081] ω j+1 =ω j +ω i (1)

[0082] Among them, ω j+1 represents the parameters of the first intermediate model obtained by the temporary server j in the current round of aggregation, ω j represents the parameters of the first intermediate model obtained after the last round of aggregation of temporary server j, ω i Represents the current local model of home terminal i.

[0083] S210: Repeat steps S207 to S209 until the termination condition is met, and the temporary server transmits the first intermediate model to the communication base station; the termination condition may be reaching a set number of iterations.

[0084] S211: The communication base station performs weighted aggregation on all received first intermediate models to obtain a second intermediate model, and returns the second intermediate model to the cloud server;

[0085] For each communication base station, all parameters of the first intermediate model are added together to obtain ω s According to the total number of temporary servers participating in the training in the communication base station, s Iterative averaging is performed, and the iterative averaging result is used as the parameter of the second intermediate model of the communication base station.

[0086] S212: The cloud server aggregates all received second intermediate models to obtain a final global model.

[0087] The cloud server adds up all the received second intermediate models to obtain ω T , and then according to the total number of communication base stations participating in the training, ω T Perform iterative averaging, and the iterative averaging result is the final global model parameter.

[0088] Example 3

[0089] In order to verify the effectiveness of the method of the present invention, the present invention also provides the following experiments:

[0090] A public real data set of home network devices is used. This data set is a statistical TCP throughput collected from a grid of home unit wireless networks placed in three different houses. The network access situation includes network access ID, type, throughput, time, number of home rooms and other information. The present invention selects the TCP throughput to collect data from a home temporary server and completes the local training task in the home node. The models trained by multiple temporary servers are updated and communicated with the base station again to perform model averaging to complete the second training. Finally, the models are aggregated by multiple base stations to complete the entire training process and obtain the final training results. The experiment was run and verified on a machine with a configuration of Intel Core i7-1165G7@2.80GHz, NVIDIA GeForceMX450 GPU and 16GB RAM. In addition, during the local training process, a variety of local training methods were used for comparison ( Figure 3 、 Figure 4 and Figure 5 "Fedfurniture" in this article represents the method of this invention. This paper also sets the same local training capacity and compares the training results with different numbers of participating base stations. When performing local training, each iteration takes 0.5 seconds, the basic communication time constant is 0.02 seconds, and the average of 5 local iterations achieves the best results. Finally, it is found that the efficiency of training with a single base station is significantly lower than that of training with multiple base stations. The comparison method uses the FedSGD algorithm and the FedAvg algorithm.

[0091] In the privacy and security experiment, the results are as follows Figure 3 As shown in the figure, federated learning has a strong track record in protecting data privacy, reducing the risk of data leakage during data transmission. Furthermore, as the number of smart home clients participating in training increases, the security of federated learning data further increases. This demonstrates that federated learning significantly protects smart home data privacy.

[0092] The learning efficiency of this method was evaluated by using different numbers of communication base station facilities. The results are as follows: Figure 4 The learning efficiency of this method is significantly improved compared to traditional federated learning, and with the increase of communication base station facilities, the overall learning efficiency will be further improved.

[0093] Finally, the present invention also evaluates the accuracy of the algorithm's predictions, such as Figure 5 As shown in the figure, in this experiment, the next throughput of TCP is predicted. The results show that this method can predict the results more accurately than traditional federated learning, and its accuracy is more adaptable to complex smart home scenarios than FedSGD and FedAVG.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An efficient federated learning method based on cloud-edge in smart home scenarios, characterized by: include: Step 1: The cloud server obtains the learning task and initializes the initial model; Step 2: The cloud server sends the learning task and initial model to the communication base stations participating in the training; Step 3: The communication base station sends the acquired learning tasks and initial models to the temporary servers and home terminals in the home units participating in the training within its communication range; The temporary server is elected by all home terminals in a family unit; The temporary server selection includes: selecting a temporary server based on the computing power and idle time of all home terminals in the home unit, specifically including: calculating the sum of the computing power and idle time of each home terminal; selecting the home terminal with the largest sum of computing power and idle time as the temporary server of the home unit; Step 4: The home terminal performs local model training and updates based on local data and the received initial model or the first intermediate model of the previous round to obtain a new local model, and transmits the local model to the temporary server; Step 5: The temporary server performs weighted aggregation on all received local models and its own local model to obtain a first intermediate model, and returns the first intermediate model to each home terminal; Step 6: Repeat steps 4 to 5 until the end condition is met, and the temporary server transmits the first intermediate model to the communication base station; Step 7: The communication base station performs weighted aggregation on all received first intermediate models to obtain a second intermediate model, and returns the second intermediate model to the cloud server; Step 8: The cloud server aggregates all received second intermediate models to obtain the final global model.

2. The efficient cloud-edge-device federated learning method in the smart home scenario according to claim 1 is characterized in that: Also includes: The temporary server collects local data from home terminals without computing power in the home unit, and combines its own local data with the received initial model or the first intermediate model of the previous round to perform local model training and update to obtain its own new local model.

3. The efficient cloud-edge-device federated learning method in the smart home scenario according to claim 1 is characterized in that: Also include before step 2: For the communication base stations that need to participate in the training, the communication base stations are registered with the cloud server, so that the cloud server can collect statistics on all the communication base stations and form a base station index; Correspondingly, step 2 is specifically as follows: the cloud server sends the learning task and the initial model to the corresponding communication base station according to the base station index.

4. The efficient cloud-edge-device federated learning method in the smart home scenario according to claim 1 is characterized in that: Also include before step 3: For the home unit that needs to participate in the training, the temporary server and / or home terminal of the home unit is registered with its corresponding communication base station, so that the communication base station can collect statistics on all temporary servers and / or home terminals and form a temporary server index; Correspondingly, step 3 is specifically as follows: the communication base station sends the acquired learning tasks and initial models to the corresponding temporary server and home terminal according to the temporary server index.

5. The efficient cloud-edge-device federated learning method in the smart home scenario according to claim 1 is characterized in that: Step 4 specifically includes: Home terminal i calculates the gradient descent of the tth iteration process based on local data and the received initial model or the first intermediate model of the previous round Where f(ω)=l(ω,D), f(ω) represents the prediction loss function, ω represents the parameters of the initial model received by home terminal i or the first intermediate model of the previous round, and D represents the local data of home terminal i. represents the gradient operator; Home terminal i descends according to the gradient Update the local model of the tth iteration process Among them, η i For the learning efficiency of home terminal i, Represents the parameters of the local model of home terminal i before the local training of the t-th iteration process.

6. The efficient cloud-edge-device federated learning method in the smart home scenario according to claim 1 is characterized in that: Step 5 specifically includes: The temporary server performs weighted aggregation on each local model according to formula (1) to obtain the parameters of the first intermediate model: oh j+1 =ω j +oh i (1) Among them, ω j+1 represents the parameters of the first intermediate model obtained by the temporary server j in the current round of aggregation, ω j represents the parameters of the first intermediate model obtained after the last round of aggregation of temporary server j, ω i Represents the current local model of home terminal i.

7. The efficient cloud-edge-device federated learning method in the smart home scenario according to claim 1 is characterized in that: Step 7 specifically includes: For each communication base station, all parameters of the first intermediate model are added together to obtain ω s According to the total number of temporary servers participating in the training in the communication base station, s Perform iterative averaging, and use the iterative averaging results as the parameters of the second intermediate model.

8. The efficient cloud-edge-device federated learning method in the smart home scenario according to claim 1 is characterized in that: The local data of the home terminal includes at least one of network throughput, power usage and camera monitoring data.

Images