Federal lightweight end-side cloud network cooperative reasoning method and system
Through the federated lightweight end-edge cloud network collaborative reasoning method of parameter pruning and device training in the cloud, the DNN model's privacy leakage and communication overhead in cross-device and cross-platform deployment and update are solved, and the DNN model is highly efficient. Low latency inference operation and model optimization.
Patent Information
- Application Number
- CN202411361819.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-27
- Publication Date
- 2025-01-07
AI Technical Summary
During the cross-device and cross-platform deployment and update of deep neural network (DNN) models, traditional methods have problems such as data privacy leakage, large communication overhead, resource limitations and insufficient model performance, especially in edge computing and IoT environments, which are difficult to Achieve efficient and low-latency inference operations.
By performing parameter pruning on cloud servers based on parameter sensitivity and significance, a lightweight model network is generated, and training is performed on the device, top-N parameters are obtained and uploaded to the cloud. Cloud servers aggregate these parameters to generate federated models. Implement cross-device deployment and update of models while reducing communication overhead and retaining local information.
It realizes efficient operation of the DNN model in resource-constrained environments, protects data privacy, reduces communication overhead, improves the efficiency and effectiveness of federated learning, adapts to the needs of different device resources, and ensures the continuous optimization and adaptability of the model.
Smart Images
Figure CN119276898A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of mobile communication technologies, and in particular to a federated lightweight end-edge-cloud network collaborative reasoning method, system, device and storage medium. Background Art
[0002] With the rapid development of edge computing and IoT technologies, more and more smart devices and terminal nodes are involved in data processing and computing tasks. The need for hierarchical deployment and collaborative reasoning of deep neural network (DNN) models in the cloud, edge servers, and terminal devices stems from the differences in resource limitations and application requirements at each computing level. The cloud has powerful computing power and storage resources, which is suitable for processing large and complex models, but there are challenges in real-time and privacy protection; the edge server is located close to the data source and can handle more complex computing tasks while reducing transmission delays and improving response speeds, but its resources are more limited than the cloud; terminal devices such as smartphones and IoT devices have the most limited computing and storage capabilities, and require extremely lightweight models to achieve low-latency, low-energy reasoning operations.
[0003] In this context, it becomes particularly important to design lightweight end-edge-cloud network collaborative optimization methods. Distributed heterogeneous devices at the end, edge, and cloud levels have differences in computing resources and storage capacity, which makes it a challenge to efficiently deploy and update DNN models in cross-device and cross-platform scenarios. Traditional centralized model training methods have problems of data privacy leakage and high communication overhead, and are not suitable for scenarios with distributed heterogeneous devices. Traditional model pruning methods may not be able to fully consider the importance of parameters and the overall performance of the model. Existing federated learning methods also face the challenge of local information loss during updates. Summary of the invention
[0004] The purpose of the present invention is to solve the shortcomings of the prior art and provide a federated lightweight end-edge-cloud network collaborative reasoning method, comprising the following steps:
[0005] The cloud server deploys the original DNN model, and performs parameter pruning on the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network, and the cloud server sends the lightweight model network to the device;
[0006] The device performs training according to the lightweight model network to obtain top-N parameters, and uploads the top-N parameters to the cloud server, and the cloud server aggregates the top-N parameters to obtain a federated model.
[0007] Preferably, the method of performing parameter pruning on the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network comprises:
[0008] Calculating a trimming index based on the original DNN model, the L1 norm, the parameter sensitivity, and the parameter significance;
[0009] The model parameters calculate parameter importance scores according to the trimming indicators, and are sorted according to the parameter importance scores;
[0010] The model parameters are set to 0 according to the final expected sparsity k of the lightweight model network, thereby obtaining the lightweight model network.
[0011] Preferably, calculating the trimming index according to the original DNN model, the L1 norm, the parameter sensitivity, and the parameter significance further includes:
[0012] The parameter θ in the original DNN model f(θ; x) with input x is replaced by a lightweight parameter θ⊙m to obtain a first DNN model f(θ⊙m; x), where m∈{0,1} d is a binary parameter mask, ⊙ is the element-by-element product calculation;
[0013] Calculating the trimming index according to the first DNN model f(θ⊙m;x), the L1 norm, the parameter sensitivity, and the parameter significance;
[0014] Among them, the cutting index S j The calculation formula of (θ; B) is as follows:
[0015]
[0016] Among them, S j represents the pruning index of the jth parameter, B is a set of samples in the training set, and L is the loss function of the parameter pruning.
[0017] Preferably, the model parameter calculates the parameter importance score according to the clipping index, further comprising:
[0018] The importance score z j By the normalized clipping index S j It turns out that:
[0019]
[0020] Where d is the number of original DNN model parameters, S k represents the kth original DNN model parameter.
[0021] Preferably, the device performs training according to the lightweight model network to obtain top-N parameters, and uploads the top-N parameters to the cloud server, further comprising:
[0022] Inputting the initialized lightweight model network parameters and the total number of training rounds to the lightweight model network;
[0023] Initialize the current training times to 0 and enter model training;
[0024] Training the lightweight model network through the local data of the device and obtaining new lightweight model parameters, obtaining a second residual model and the top-N parameters according to the new lightweight model parameters, and sending the top-N parameters to the cloud server to complete the model training;
[0025] If the current number of training times does not reach the total number of training rounds, the current number of training times is increased by 1 and the model training is entered again. If the current number of training times reaches the total number of training rounds, the training is completed and a federated model is obtained.
[0026] Preferably, obtaining the second residual model and top-N parameters according to the new lightweight model parameters further includes:
[0027] adding the new lightweight model parameters to the first residual model;
[0028] Filtering the first residual model to obtain the top-N parameters;
[0029] The second residual model is obtained by subtracting the top-N parameters in the first residual model, and the top-N parameters are uploaded to the cloud server.
[0030] Preferably, in step S3, the cloud server aggregates the top-N parameters and obtains a federated model, further comprising:
[0031] The cloud server receives the top-N parameters, calculates a weighted average of the top-N parameters to obtain a federated model, and sends the federated model to the device;
[0032] The federation model W t+1 The calculation formula is as follows:
[0033] W t+1 =FedAvg(msg),
[0034] Among them, msg is the top-N parameter.
[0035] Based on the same concept, the present invention also provides a federated lightweight end-edge-cloud network collaborative reasoning system, which is characterized by comprising:
[0036] A pruning module is used for deploying the original DNN model on a cloud server, and pruning the parameters of the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network, and the cloud server sends the lightweight model network to the device;
[0037] A federation module is used for the device to train the lightweight model network to obtain top-N parameters, and upload the top-N parameters to the cloud server, and the cloud server aggregates the top-N parameters to obtain a federated model.
[0038] Based on the same concept, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the federated lightweight end-edge cloud network collaborative reasoning method as described in any one of the embodiments.
[0039] Based on the same concept, the present invention also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the federated lightweight end-edge-cloud network collaborative reasoning method as described in any one of the embodiments.
[0040] Compared with the prior art, the present invention has the following beneficial effects:
[0041] 1) The present invention compresses the DNN model into a lightweight model suitable for heterogeneous devices with different levels of resources and requirements by performing parameter pruning of the DNN model to different degrees, and deploys the model across devices;
[0042] 2) The present invention performs parameter pruning based on parameter sensitivity and parameter significance to accurately remove redundant parameters, ensuring that the pruned model can run efficiently in a resource-constrained environment while maintaining the overall performance of the model;
[0043] 3) The present invention obtains top-N parameters by training the model on the device, and uploads the top-N parameters to the cloud server. The cloud server aggregates the top-N parameters to obtain a federated model, thereby solving the problem of federal updating of the local model during system operation, protecting data privacy, and further realizing continuous optimization and adaptability of the model.
[0044] 4) The present invention obtains a federated model by implementing a Top-N sparse residual model, thereby achieving sparse processing of model parameters, reducing communication overhead, retaining local model information, and improving the efficiency and effect of the federated learning process. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.
[0046] Figure 1 A flowchart of a federated lightweight end-edge-cloud network collaborative reasoning method of the present invention;
[0047] Figure 2 This is an architecture diagram of a federated lightweight end-edge-cloud network collaborative reasoning method of the present invention;
[0048] Figure 3 A training diagram of a federated lightweight end-edge-cloud network collaborative reasoning method of the present invention. DETAILED DESCRIPTION
[0049] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.
[0050] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.
[0051] The technical terms involved in the embodiments of the present invention are defined as follows:
[0052] 1. A deep neural network (DNN) model is a computational process that uses a trained neural network to predict or classify new data. The model consists of multiple layers of neurons connected by nonlinear activation functions to form a complex structure for simulating and processing complex input data. During the inference process, the model receives input data, calculates through each layer, and finally generates output results. DNN inference models are widely used in image recognition, speech recognition, natural language processing and other fields, relying on their deep feature extraction and representation capabilities to achieve high-precision prediction and decision-making.
[0053] 2. Model pruning is a method for optimizing DNN models. It reduces the size and complexity of the model by removing unimportant parameters and computations to adapt to the resource constraints of edge and terminal devices. The weight pruning method removes connections with small and unimportant weights, thereby reducing the computational requirements and storage of the model; structured pruning is further optimized by removing entire filters, channels, or layers, which brings more significant acceleration effects and makes the model run more efficiently on hardware accelerators. Through these pruning techniques, while maintaining model performance, its resource consumption can be greatly reduced, thereby achieving efficient reasoning.
[0054] 3. Federated Learning provides a method for model training and optimization without sharing data. Each data holder can train the model on a local device and only transmit the model parameters to a central server instead of transmitting the original data. This distributed learning model not only protects data privacy, but also effectively reduces the amount of data transmission and communication overhead. In addition, with the development of edge computing and Internet of Things technologies, more and more data is generated from edge devices and terminal devices, and traditional centralized learning methods can no longer meet the needs of distributed environments. It provides a new solution for cross-device and cross-platform machine learning, making model training and optimization more intelligent, more secure, and with lower communication costs.
[0055] 4. Edge computing is a distributed computing architecture that reduces latency and bandwidth usage by processing data close to the data source. It moves computing resources and data storage to devices and local servers at the edge of the network, rather than relying on remote data centers. This enables faster response times and higher efficiency, and is particularly suitable for applications that require real-time processing and low latency, such as IoT devices, smart cities, self-driving cars, and industrial automation. Edge computing not only improves system performance, but also enhances data privacy and security because data does not have to leave the local network for processing.
[0056] Example 1
[0057] See also Figure 1 As shown, a federated lightweight end-edge-cloud network collaborative reasoning method provided in this embodiment includes the following steps:
[0058] The cloud server deploys the original DNN model, and performs parameter pruning on the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network, and the cloud server sends the lightweight model network to the device. Specifically, in this embodiment, the original DNN model is parameter pruned according to parameter sensitivity and parameter significance to obtain an intermediate model network, and the cloud server sends the intermediate model network to the edge server to compress the original DNN model into different lightweight versions suitable for the cloud, edge server and terminal devices.
[0059] By performing hierarchical lightweight optimization at different levels, the DNN model can perform complex calculations in the cloud, achieve medium-scale computing loads on edge servers, and achieve efficient and real-time reasoning on terminal devices, meeting the performance and resource requirements of each level.
[0060] Preferably, the method of performing parameter pruning on the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network comprises:
[0061] Calculating a trimming index based on the original DNN model, the L1 norm, the parameter sensitivity, and the parameter significance;
[0062] The model parameters calculate parameter importance scores according to the trimming indicators, and are sorted according to the parameter importance scores;
[0063] The model parameters are set to 0 according to the final expected sparsity k of the lightweight model network, and the lightweight model network is obtained. Specifically, in this embodiment, the ultimate model network is obtained by pruning through the final sparsity of the intermediate model network.
[0064] Preferably, calculating the trimming index according to the original DNN model, the L1 norm, the parameter sensitivity, and the parameter significance further includes:
[0065] The parameter θ in the original DNN model f(θ; x) with input x is replaced by a lightweight parameter θ⊙m to obtain a first DNN model f(θ⊙m; x), where m∈{0,1} d is a binary parameter mask, ⊙ is the element-by-element product calculation;
[0066] Calculating the trimming index according to the first DNN model f(θ⊙m;x), the L1 norm, the parameter sensitivity, and the parameter significance;
[0067] Among them, the cutting index S j The calculation formula of (θ; B) is as follows:
[0068]
[0069] Among them, S j represents the pruning index of the jth parameter, B is a set of samples in the training set, and L is the loss function of the parameter pruning.
[0070] Preferably, the model parameter calculates the parameter importance score according to the clipping index, further comprising:
[0071] The importance score z j By the normalized clipping index S j It turns out that:
[0072]
[0073] Where d is the number of original DNN model parameters, S k represents the kth original DNN model parameter.
[0074] Preferably, unimportant connections in the model are removed through weight pruning technology, thereby reducing the number of model parameters and the amount of computation, making it more efficient to run on hardware devices. The pruned model is allocated according to the computing power and storage resources of each level of equipment: the cloud deploys complex complete models, edge servers deploy models of medium complexity, and terminal devices deploy the lightest models. This hierarchical pruning and deployment strategy ensures that different devices can run efficiently within their respective resource constraints.
[0075] The device trains the lightweight model network to obtain top-N parameters, and uploads the top-N parameters to the cloud server. The cloud server aggregates the top-N parameters and obtains the federated model. Specifically, in this embodiment, the device trains the intermediate network model to obtain top-N parameters, and uploads the top-N parameters to the cloud server.
[0076] Preferably, the device trains the lightweight model network to obtain top-N parameters, and uploads the top-N parameters to the cloud server, further comprising:
[0077] Input the initialized lightweight model network parameters W and the total training rounds T to the lightweight model network;
[0078] Initialize the current training times to 0 and enter model training. Specifically, in this embodiment, for the current training times t=1, 2, ..., T, iterate in a loop;
[0079] The lightweight model network is trained by the local data of the device, and new lightweight model parameters are obtained. The second residual model and top-N parameters are obtained according to the new lightweight model parameters, and the top-N parameters are sent to the cloud server to complete the model training. Specifically, in this embodiment, the lightweight model network is trained for the local data of each device or edge i;
[0080] If the current number of training times does not reach the total number of training rounds, the current number of training times is increased by 1 and the model training is started again. If the current number of training times reaches the total number of training rounds, the training is completed and the federated model is obtained.
[0081] Each device uses its own data to train the model locally. After completing a round of local training, the updated model parameters are sent to the central server (cloud server). The central server aggregates the model updates from all devices, generates a global model, and sends it to each device for the next round of local training. This process ensures that each device can achieve overall optimization and update of the model through federated learning without sharing original data. In order to reduce data transmission during the model update process.
[0082] Preferably, obtaining the second residual model and top-N parameters according to the new lightweight model parameters further includes:
[0083] The new lightweight model parameters Add to the first residual model Specifically, the formula of the first residual model is
[0084] Filter the first residual model to obtain top-N parameters. Specifically, in this embodiment, top-N parameters The calculation formula is
[0085] Subtract the top-N parameters in the first residual model to obtain a second residual model, and upload the top-N parameters to the cloud server. Specifically, in this embodiment, the second residual model The calculation formula is:
[0086] Preferably, in step S3, the cloud server aggregates the top-N parameters and obtains a federated model, further comprising:
[0087] The cloud server receives the top-N parameters, calculates the weighted average of the top-N parameters to obtain the federated model, and sends the federated model to the device;
[0088] Federal Model W t+1 The calculation formula is as follows:
[0089] W t+1=FedA vg(msg),
[0090] Among them, msg is the top-N parameter.
[0091] Preferably, device i selects the top-N parameters with the largest update amplitude in the tth round of training and sends them to the server. The remaining unselected parameter set is called the residual model RM t , the residual model is retained locally, and when the device participates in the t+1th round of FL training, the residual model will be added to the next local learning to recover the lost features.
[0092] Example 2
[0093] See also Figure 2 and Figure 3 As shown, this embodiment also provides a federated lightweight end-edge-cloud network collaborative reasoning system, which is characterized by including:
[0094] A pruning module is used for deploying the original DNN model on a cloud server, and pruning the parameters of the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network, and the cloud server sends the lightweight model network to the device;
[0095] A federation module is used for the device to train according to the lightweight model network to obtain top-N parameters, and upload the top-N parameters to the cloud server, and the cloud server aggregates the top-N parameters to obtain a federated model. Specifically, in this embodiment, the model federation training module is used to perform federal updates on the local model during system operation, protect data privacy, reduce transmission overhead, and utilize the computing resources of devices at all levels to achieve distributed collaborative optimization.
[0096] Example 3
[0097] In some embodiments of the present application, a computer device is also provided, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor performs the steps of the federated lightweight end-edge cloud network collaborative reasoning method as described in any one of Example 1.
[0098] The present invention also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the federated lightweight end-edge-cloud network collaborative reasoning method as described in any one of Example 1.
[0099] It can be understood that for the aforementioned federated lightweight end-edge cloud network collaborative reasoning method, if it is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.
[0100] Computer readable storage media may include data signals propagated in baseband or as part of a carrier wave, wherein readable program codes are carried. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program codes contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination thereof.
[0101] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. A federated lightweight end-edge-cloud network collaborative reasoning method, characterized in that: The following steps are involved: The cloud server deploys the original DNN model, and performs parameter pruning on the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network, and the cloud server sends the lightweight model network to the device; The device performs training according to the lightweight model network to obtain top-N parameters, and uploads the top-N parameters to the cloud server, and the cloud server aggregates the top-N parameters to obtain a federated model.
2. According to claim 1, a federated lightweight end-edge-cloud network collaborative reasoning method is characterized in that: The method of performing parameter pruning on the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network includes: Calculating a trimming index based on the original DNN model, the L1 norm, the parameter sensitivity, and the parameter significance; The model parameters calculate parameter importance scores according to the trimming indicators, and are sorted according to the parameter importance scores; The model parameters are set to 0 according to the final expected sparsity k of the lightweight model network, thereby obtaining the lightweight model network.
3. According to claim 2, a federated lightweight end-edge-cloud network collaborative reasoning method is characterized in that: Calculating the trimming index according to the original DNN model, the L1 norm, the parameter sensitivity, and the parameter significance, further includes: The parameter θ in the original DNN model f(θ; x) with input x is replaced by a lightweight parameter θ⊙m to obtain a first DNN model f(θ⊙m; x), where m∈{0,1} d is a binary parameter mask, ⊙ is the element-by-element product calculation; Calculating the trimming index according to the first DNN model f(θ⊙m;x), the L1 norm, the parameter sensitivity, and the parameter significance; Among them, the cutting index S j The calculation formula of (θ; B) is as follows: Among them, S j represents the pruning index of the jth parameter, B is a set of samples in the training set, and L is the loss function of the parameter pruning.
4. According to claim 3, a federated lightweight end-edge-cloud network collaborative reasoning method is characterized in that: The model parameter calculates the parameter importance score according to the clipping index, further comprising: The importance score z j By the normalized clipping index S j It turns out that: Where d is the number of original DNN model parameters, S k represents the kth original DNN model parameter.
5. According to claim 4, a federated lightweight end-edge-cloud network collaborative reasoning method is characterized in that: The device performs training according to the lightweight model network to obtain top-N parameters, and uploads the top-N parameters to the cloud server, further comprising: Inputting the initialized lightweight model network parameters and the total number of training rounds to the lightweight model network; Initialize the current training times to 0 and enter model training; Training the lightweight model network through the local data of the device and obtaining new lightweight model parameters, obtaining a second residual model and the top-N parameters according to the new lightweight model parameters, and sending the top-N parameters to the cloud server to complete the model training; If the current number of training times does not reach the total number of training rounds, the current number of training times is increased by 1 and the model training is entered again. If the current number of training times reaches the total number of training rounds, the training is completed and a federated model is obtained.
6. A federated lightweight end-edge-cloud network collaborative reasoning method according to claim 5, characterized in that: The obtaining of a second residual model and top-N parameters according to the new lightweight model parameters further includes: adding the new lightweight model parameters to the first residual model; Filtering the first residual model to obtain the top-N parameters; The second residual model is obtained by subtracting the top-N parameters in the first residual model, and the top-N parameters are uploaded to the cloud server.
7. A federated lightweight end-edge-cloud network collaborative reasoning method according to claim 6, characterized in that: In step S3, the cloud server aggregates the top-N parameters and obtains a federated model, further comprising: The cloud server receives the top-N parameters, calculates a weighted average of the top-N parameters to obtain a federated model, and sends the federated model to the device; The federation model W t+1 The calculation formula is as follows: W t+1 =FedAvg(msg), Among them, msg is the top-N parameter.
8. A federated lightweight end-edge-cloud network collaborative reasoning system, characterized in that: include: A pruning module is used for deploying the original DNN model on a cloud server, and pruning the parameters of the original DNN model according to parameter sensitivity and parameter significance to obtain a lightweight model network, and the cloud server sends the lightweight model network to the device; A federation module is used for the device to train the lightweight model network to obtain top-N parameters, and upload the top-N parameters to the cloud server, and the cloud server aggregates the top-N parameters to obtain a federated model.
9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the federated lightweight end-edge-cloud network collaborative reasoning method as described in any one of claims 1 to 8.
10. A storage medium storing computer-readable instructions, characterized in that: When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the federated lightweight end-edge-cloud network collaborative reasoning method as described in any one of claims 1 to 8.