Efficient joint processing method and device under computing power network

Abstract

The invention provides an efficient joint processing method and device under a computing power network, and the method comprises the steps: a federated learning server selects a plurality of heterogeneous terminals in the computing power network, and transmits the heterogeneous terminals to each heterogeneous terminal; the edge server carries out model training on the initial model by using local data of the corresponding heterogeneous terminal to obtain local model parameters; performing model adjustment through an adjustment training strategy of the soft mask strategy gradient; the edge server performs heterogeneous data processing on the generated local model parameters through a reweighted adversarial graph convolutional network to obtain a model training result of the corresponding heterogeneous terminal; and the cloud server aggregates the model training results of the plurality of edge servers and issues the aggregated global model to the edge servers, and the edge servers forward the aggregated global model to the corresponding heterogeneous terminals, so that the corresponding heterogeneous terminals update local model parameters according to the global model. According to the method, efficient joint calculation of the heterogeneous terminals under the computing power network is realized, and the model accuracy and the training efficiency are improved.

Description

A method and apparatus for efficient joint processing in a computing power network Technical Field

[0001] This application relates to the field of computing power network technology, and more specifically, to an efficient joint processing method and apparatus under a computing power network. Background Technology

[0002] With the development of 5G, artificial intelligence, and edge computing technologies, a large number of applications with high bandwidth, low latency, and high computing power requirements have emerged in vertical industries, such as artificial intelligence and multimedia rendering applications. This has led to an explosive growth in the number of connections and data volume, creating an urgent demand for computing power. Computing power networks aim to interconnect and coordinate distributed computing nodes, achieving efficient utilization of network and computing resources through improvements in network architecture and protocols. Federated learning, as a collaborative computing model, is widely used in computing power networks. It allows terminals to train local models using local data and then upload the model parameters to a central server for aggregation, thereby achieving global model co-construction while ensuring data privacy.

[0003] However, existing federated learning technologies face challenges from heterogeneous terminal devices (such as differences in computing power and communication resources) and data heterogeneity (such as non-independent and identically distributed data). Device heterogeneity makes it difficult for some low-computing devices to participate in training or results in slow model convergence; data heterogeneity leads to global model bias and performance degradation. Existing research attempts to alleviate device heterogeneity through methods such as client selection, model compression, and adaptive aggregation, but these methods suffer from low participation of low-computing devices and poor model fairness. For data heterogeneity, methods such as data augmentation and knowledge distillation are used, but these methods still have drawbacks such as high communication overhead, long training time, and privacy risks.

[0004] It is evident that existing technologies, due to the diversity of terminal types and the heterogeneity of computing and data resources in computing networks, face difficulties in terminal computing collaboration, resulting in poor global model convergence, low training efficiency, and an inability to achieve efficient allocation of computing network resources while ensuring privacy. Summary of the Invention

[0005] The purpose of this application is to provide an efficient joint processing method and apparatus under a computing power network, which can achieve heterogeneous data collaboration by adaptively adjusting the training strategy through a soft masking strategy gradient and a reweighted adversarial graph convolutional network, thereby realizing efficient joint computing of heterogeneous terminals under a computing power network, improving model accuracy and training efficiency, while ensuring data privacy.

[0006] Firstly, a highly efficient joint processing method under a computing power network is provided, applicable to a highly efficient joint processing system including a federated learning server, edge servers, and a cloud server. This method may include: the federated learning server selecting multiple heterogeneous terminals in the computing power network and distributing an initial model to each heterogeneous terminal; the edge server using local data from the corresponding heterogeneous terminal to train the initial model, obtaining the trained local model parameters; wherein, model adjustment is performed using a soft masking strategy gradient adjustment training strategy, the adjustment training strategy including modeling the computational and communication energy consumption of the corresponding heterogeneous terminal; the edge server processing the generated local model parameters through a reweighted adversarial graph convolutional network to obtain the model training results for the corresponding heterogeneous terminal; the cloud server aggregating the model training results from multiple edge servers and distributing the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals, enabling the corresponding heterogeneous terminals to update their local model parameters according to the global model.

[0007] In one possible implementation, the federated learning server selects multiple heterogeneous terminals from the computing power network through a programmatic algorithm, including: selecting heterogeneous terminals to participate in training from the terminal set using a random selection algorithm or a greedy algorithm based on the terminal's computing power, communication status, or resource availability indicators; and distributing the initial model parameters to the selected heterogeneous terminals in the form of a weight matrix via a network protocol.

[0008] In one possible implementation, modeling the computational and communication energy consumption of the corresponding heterogeneous terminals includes: computational energy consumption: calculated based on the number of CPU cycles required to execute one round of training and the computing power of the heterogeneous terminals; wherein, computational energy consumption is directly proportional to the number of CPU cycles and inversely proportional to the computing power; communication energy consumption: calculated based on the size of the neural network model, the uplink sub-channel bandwidth, and the signal-to-noise ratio of the heterogeneous terminals; wherein, communication energy consumption is directly proportional to the model size and inversely proportional to the channel capacity.

[0009] In one possible implementation, the training strategy for adjusting the soft mask policy gradient also includes: defining the state space as consisting of training states, available computing resources, optimization metrics, and time delay; defining the action space as consisting of the number of local training iterations on the terminal; and constructing a reward function based on a weighted calculation of model accuracy, energy consumption, and time delay.

[0010] In one possible implementation, heterogeneous data is processed using a reweighted adversarial graph convolutional network, including: using the weighted network to output sample weights, the classifier to output prediction results, and adversarial training by minimizing the weighted cross-entropy loss and maximizing the regularization term; introducing a feature discriminator to extract feature representations from the local model and outputting feature prediction results through the Softmax function, calculating the unified adversarial loss based on the output probability of the feature discriminator; and constructing a group consensus loss and a global consensus loss, where the group consensus loss is based on KL divergence to calculate the difference between local predictions and group consensus predictions, and the global consensus loss is based on KL divergence to calculate the difference between local predictions and global model predictions.

[0011] In one possible implementation, the formula for calculating the unified adversarial loss is expressed as: ,in The output probability of the feature discriminator for sample i on the j-th terminal model is calculated using a logarithmic function and average aggregation.

[0012] In one possible implementation, the cloud server aggregates the model training results from multiple edge servers and distributes the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals. This includes: aggregating the model parameters uploaded by the edge servers using a weighted average algorithm; encoding the aggregated global model parameters into a binary stream and distributing it to the edge servers via a network protocol; the edge servers parsing and forwarding the stream to the corresponding heterogeneous terminals, so that the heterogeneous terminals can update their models by overwriting the locally stored model files after receiving the global model parameters.

[0013] Secondly, a high-efficiency joint processing device under a computing power network is provided, which is applied in a high-efficiency joint processing system including federated learning servers, edge servers, and cloud servers. The device may include: a sending unit, used to select multiple heterogeneous terminals in the computing power network and send an initial model to each heterogeneous terminal; a training unit, used to train the initial model using local data of the corresponding heterogeneous terminals to obtain the trained local model parameters; wherein, the model is adjusted by a training strategy of adjusting the gradient of a soft masking policy, and the adjustment training strategy includes modeling the computing energy consumption and communication energy consumption of the corresponding heterogeneous terminals; a processing unit, used to process the generated local model parameters through a reweighted adversarial graph convolutional network to obtain the model training results of the corresponding heterogeneous terminals; and an aggregation unit, used to aggregate the model training results of multiple edge servers and send the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals so that the corresponding heterogeneous terminals update their local model parameters according to the global model.

[0014] Thirdly, an electronic device is provided, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory is used to store computer programs; and the processor, when executing the program stored in the memory, implements any of the steps described in the first aspect above.

[0015] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the steps of any of the methods described in the first aspect above.

[0016] This application provides an efficient joint processing method and apparatus under a computing power network. The method is applied to an efficient joint processing system including a federated learning server, edge servers, and a cloud server. The method includes: the federated learning server selecting multiple heterogeneous terminals in the computing power network and distributing an initial model to each heterogeneous terminal; the edge server using local data from the corresponding heterogeneous terminal to train the initial model, obtaining the trained local model parameters; adjusting the model through a soft masking strategy gradient adjustment training strategy, which includes modeling the computational and communication energy consumption of the corresponding heterogeneous terminals; the edge server processing the generated local model parameters through a reweighted adversarial graph convolutional network to obtain the model training results for the corresponding heterogeneous terminals; and the cloud server aggregating the model training results from multiple edge servers and distributing the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals, enabling the corresponding heterogeneous terminals to update their local model parameters according to the global model. This method achieves efficient joint computation of heterogeneous terminals under a computing power network, improving model accuracy and training efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 is a flowchart illustrating an efficient joint processing method under a computing network according to an embodiment of this application; Figure 2 is a structural diagram illustrating an efficient joint processing device under a computing network according to an embodiment of this application; Figure 3 is a structural diagram illustrating an electronic device according to an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art. The words "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are only used to distinguish different components. The words "comprising" or "including," etc., mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but do not exclude other elements or objects. The words "connected," "coupled," or "connected," etc., are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] The efficient joint processing method under a computing power network provided in this application embodiment can be applied to a system architecture of federated learning servers, heterogeneous terminals (such as mobile devices and IoT nodes), edge servers, and cloud servers. The federated learning server is deployed in the cloud and is responsible for global task scheduling; the edge server is deployed at the network edge (such as at a base station) and is responsible for regional model aggregation; heterogeneous terminals have local computing and storage capabilities and connect to the edge server via a wireless network (such as 5G). During initialization, the federated learning server selects multiple heterogeneous terminals from the computing power network using a selection algorithm (such as random selection or a greedy algorithm based on resource status). The selection criteria include the terminal's computing power, memory size, or network connection status. The server sends the initial model parameters (in the format of a neural network weight matrix, such as an HDF5 file) to the selected terminals via a secure communication protocol (such as HTTPS).

[0021] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0022] Figure 1 is a flowchart illustrating an efficient federated processing method under a computing power network according to an embodiment of this application. As shown in Figure 1, the method may include: step S110, whereby the federated learning server selects multiple heterogeneous terminals in the computing power network and distributes the initial model to each heterogeneous terminal.

[0023] This step is the initialization phase, designed to select suitable heterogeneous terminals from the computing power network to participate in the federated learning task. The computing power network consists of multiple terminal devices (such as mobile phones, IoT devices, edge servers, etc.), which vary significantly in terms of computing power, communication status, and resource availability. The federated learning server (usually deployed in the cloud) is responsible for global coordination.

[0024] In practice, based on the terminal's computing power, communication status, or resource availability indicators, a random selection algorithm or a greedy algorithm is used to select heterogeneous terminals from the terminal set to participate in training; the initial model parameters are distributed to the selected heterogeneous terminals in the form of a weight matrix via a network protocol.

[0025] Specifically, the federated learning server uses a programmed algorithm to select heterogeneous terminals from the terminal set to participate in training. Selection criteria are based on the terminal's computing power (e.g., CPU clock speed), communication status (e.g., signal-to-noise ratio, SINR), or resource availability metrics (e.g., remaining battery power). In practical applications, random selection algorithms (e.g., uniform random sampling) or greedy algorithms (e.g., prioritizing terminals with high computing power) can be used for selection. For example, a greedy algorithm calculates a resource score for each terminal (e.g., a weighted sum of computing power and communication bandwidth) and selects the top K terminals with the highest scores (K being a preset value) as heterogeneous terminals. This selection process ensures terminal diversity, avoids permanently excluding low-computing-power devices, and thus improves model fairness.

[0026] For model delivery: The initial model is a pre-trained neural network model (such as a convolutional neural network), whose parameters are encoded in weight matrix format. The server delivers the initial model parameters to the selected terminal via a network protocol (such as HTTP or MQTT). The weight matrix is ​​serialized into a binary stream and transmitted securely using encryption. Upon receiving the data, the terminal parses the binary stream and loads the model into its local memory, ready for training.

[0027] This step lays the foundation for subsequent training by using smart terminals to balance resource utilization and model performance.

[0028] Step S120: The edge server uses the local data of the corresponding heterogeneous terminal to train the initial model and obtain the local model parameters generated by the training.

[0029] This step is the core training phase, where edge servers (such as base stations or edge computing nodes) coordinate local model training on their subordinate heterogeneous terminals. To address terminal heterogeneity, a soft-masked policy gradient is used to adaptively adjust the training strategy, optimizing the trade-off between model accuracy, energy consumption, and time latency.

[0030] First, the resource consumption of each terminal i is quantitatively modeled, including computational energy consumption, communication energy consumption, and model accuracy. Assume terminal i participates in the t-th round of federated learning training: Computational energy consumption: calculated based on the number of CPU cycles required to execute one round of training and the computing power of heterogeneous terminals; where computational energy consumption is directly proportional to the number of CPU cycles and inversely proportional to computing power; specifically: the computational energy consumption generated by terminal i completing one round of local training. The calculation formula is: , where F is the number of CPU cycles required to perform one round of training (obtained through a hardware performance counter). The computing power of the heterogeneous terminal i (e.g., CPU clock speed, in GHz). This is a normalization factor (preset constant 1) for calculating computing resource consumption. The model shows that devices with weaker computing power consume more energy.

[0031] Communication energy consumption is calculated based on the size of the neural network model, the uplink sub-channel bandwidth, and the signal-to-noise ratio of heterogeneous terminals. The communication energy consumption is directly proportional to the model size and inversely proportional to the channel capacity.

[0032] Specifically: Communication energy consumption generated by terminal i uploading local model parameters The calculation formula is: Where M is the size of the neural network model (in bits), W is the sub-channel bandwidth (read from the network interface, e.g., 10MHz), and C is the set of uplink sub-channels. It is a binary indicator (1 when terminal i uses channel c, otherwise 0). The signal-to-noise ratio of terminal i (obtained through the channel measurement module). This is a normalization factor for communication resource consumption. The model reflects the impact of channel conditions on energy consumption.

[0033] Furthermore, total energy consumption It is derived by weighted summation of computational energy consumption and communication energy consumption, and is used for subsequent resource constraint optimization.

[0034]

[0035] in, This is a key decision variable, representing the number of local iterations performed by terminal i in round t. Model accuracy. Test dataset on the server side Above assessment:

[0036] in, It is the loss function (1 if the prediction is correct, 0 otherwise). If and only if hour, , It is a predicted value. yes arrive The mapping function.

[0037] Due to the heterogeneity of terminals, global model aggregation is affected by terminals that upload models later. Furthermore, the communication resources, computing resources, and latency of a terminal are all related to its local training frequency. To determine the optimal local update frequency in a dynamic communication environment and achieve the best trade-off between model accuracy and time latency within a given resource budget, a common optimization objective is set for all terminals under constraints of total resources and latency. The common optimization objective for terminal i is defined as follows:

[0038]

[0039] in, Indicates the upper limit of the resource budget. It's a time delay. In the t-th round of federation, a set of frequency decision-making strategies from the previous rounds is learned. This represents the number of local iterations performed by terminal i in the t-th round of federated learning. (Using...) , Factors characterizing the terminal optimization objective. , To simplify the process, the long-term effects of the entire training process are broken down into the short-term training effects of each round of federated learning. question.

[0040] First, virtual queues are constructed to meet the constraints of the problem to be solved. A definition is defined to describe the resource consumption of all virtual queues during t rounds of training, while minimizing the resource overhead of each round of federated learning to maintain queue stability. The evolution of resource-scarce queues is shown in the following equation:

[0041] For the optimization goals of the terminal, the original problem can be... Transform into As shown in the following formula:

[0042]

[0043] In the formula Indicates terminal The set of probabilistic decision-making strategies prior to the t-th round of federated learning, where , It is the total number of global rounds in federated learning. This indicates that the accuracy threshold is a constant. . It is the total energy consumption of device i during the t-th round of federated learning training, as shown in the energy consumption calculation formula above, where .

[0044] Further details on the implementation of the soft masking policy gradient: The edge server achieves adaptive adjustment of the training policy through the following procedural operations: State space definition: The system state includes the training state. (e.g., model loss value), available computing resources (such as CPU utilization), optimization metrics (Including accuracy weights and latency weights), time latency (e.g., training time). Status information is collected by the terminal and uploaded to the edge server.

[0045] Action space definition: Action This represents the number of times terminal i has undergone local training in round t, with values ​​ranging from integers (e.g., 1-10 times). Actions are decided by the edge server based on state.

[0046] reward function Quantify the training effect of each round: ,in The model accuracy was calculated using forward inference on a test dataset. The preset threshold is used, and v and Q are positive control parameters (e.g., v=1, Q=0.5). This represents the available computing resources of terminal i, used to penalize excessive resource consumption. The reward function encourages high accuracy, low energy consumption, and low latency.

[0047] Strategy optimization: The goal is to find the optimal strategy. Maximize cumulative discount rewards: ,in This is the reward discount factor. The edge server uses a multi-agent reinforcement learning framework and adaptively updates the target network (such as Polyak averaging) through a soft masking mechanism to accelerate policy convergence. The specific process includes: Step S130, the edge server processes the generated local model parameters through a reweighted adversarial graph convolutional network to obtain the model training results for the corresponding heterogeneous terminals.

[0048] This step addresses data heterogeneity (non-independent and identically distributed, Non-IID) by using a reweighted adversarial graphical convolutional network (RA-GCN) to process local model parameters and mitigate model bias. RA-GCN ensures fairness across heterogeneous data through adversarial training and feature alignment.

[0049] In practice, the generated local model parameters are processed through a reweighted adversarial graph convolutional network to obtain the model training results for the corresponding heterogeneous terminals.

[0050] Specifically: The edge server inputs the local model parameters uploaded by the terminal into RA-GCN and performs the following operations: uses a weighted network to output sample weights, a classifier to output prediction results, and performs adversarial training by minimizing the weighted cross-entropy loss and maximizing the regularization term; introduces a feature discriminator to extract feature representations from the local model and outputs feature prediction results through the Softmax function, and calculates the unified adversarial loss based on the output probability of the feature discriminator; constructs a group consensus loss and a global consensus loss, where the group consensus loss is based on KL divergence to calculate the difference between local predictions and group consensus predictions, and the global consensus loss is based on KL divergence to calculate the difference between local predictions and global model predictions.

[0051] Endogenous Model Training and RA-GCN Application: Assume that K terminals are selected for training in each round, and the dataset of terminal k is... ,in It's about the amount of data. Model h is determined by the feature extractor. and classifier composition.

[0052] RA-GCN Training: RA-GCN is trained adversarially between a weighted network and a classifier. The loss function is:

[0053] in, These are the sample weights output by the weighted network. It is the predicted probability output by the classifier. It is a weighted network. This is the regularization coefficient (e.g., 0.1). The first part is the dynamically weighted cross-entropy loss, and the second part is the regularization term to prevent overfitting. During training, the weighted network (maximizing the loss) and the classifier (minimizing the loss) are alternately optimized to balance the class distribution.

[0054] Unified Feature Representation Learning: To align feature representations across different terminals, a feature discriminator D is introduced. Feature Extraction: For samples... The local model output features of terminal j ,in These are local parameters.

[0055] Feature identification: posterior prediction is calculated using the Softmax function. ,in These are the discriminator parameters.

[0056] Unified adversarial loss: The formula for calculating unified adversarial loss is as follows: ,in The output probability of the feature discriminator for sample i on the j-th terminal model is calculated using a logarithmic function and average aggregation. This loss enables the discriminator to distinguish features from different models and mitigates parameter differences by adjusting parameters to make the discrimination probabilities equal.

[0057] Consensus Loss Construction: To reduce inconsistencies between local and global optimization goals, we define group consensus loss and global consensus loss: Group consensus loss: In the client group Internally, calculate consensus prediction ,in, , These are weights. The loss is the KL divergence:

[0058] Global consensus loss: aligned with global model predictions.

[0059] in, .

[0060] After training the model locally, the terminal uploads the model parameters to the edge server. The server applies RA-GCN for feature alignment and loss optimization, and outputs the processed model training results.

[0061] The model is fine-tuned using unified adversarial loss, group consensus loss, and global consensus loss. Building upon this, this application further defines the loss function of the global endogenous model using a binary classification cross-entropy method, balancing global consensus loss and local group loss. Then, with the goal of minimizing the loss function, the endogenous model is iteratively optimized, effectively mitigating the bias problem of the endogenous model when dealing with heterogeneous data.

[0062] In one example, the specific implementation of the system interaction process for heterogeneous terminal joint training includes the following steps: Step 1: In the computing power network, each computing power terminal collects data based on local decisions of <state, action, reward, adjacency state>, and integrates information from the federated learning environment (terminal environment, such as channel state; computing power network environment, such as communication capability, computing capability, energy state) and transmits it to the edge server.

[0063] Step 2: The edge server performs global optimization through a soft mask gradient strategy, and jointly determines the training frequency of the computing power network using the state space, reward function, and state transition graph. The model performance is measured by training loss value and test accuracy, and finally the adaptive environment quickly converges and feeds back to the computing power terminal.

[0064] Step 3: After optimization, expand the experience buffer dimension and feed back the <action, state, reward> of the computing power node to the computing power terminal in the form of a learning strategy.

[0065] Step 4: Finally, the computing terminals in the computing power network update their status, update the federated learning environment and training objectives with the feedback obtained, and update the policy based on the optimization results to proceed with the next round of training.

[0066] Step S140: The cloud server aggregates the model training results of multiple edge servers and distributes the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals so that the corresponding heterogeneous terminals can update their local model parameters according to the global model.

[0067] This step completes global model aggregation and updates, enabling edge-cloud collaboration. The cloud server (such as a central data center) is responsible for integrating the models from all edge servers.

[0068] In practice, a weighted average algorithm is used to aggregate the model parameters uploaded by the edge servers. The aggregated global model parameters are encoded into a binary stream and sent to the edge servers via a network protocol. The edge servers parse and forward the stream to the corresponding heterogeneous terminals, allowing them to receive the global model parameters and update their models by overwriting their locally stored model files. Specifically: Model aggregation mechanism: The cloud server uses a weighted average algorithm (such as FedAvg) to aggregate the model parameters. Assume there are M edge servers, each uploading model parameters... Global model parameters after aggregation ,in These are weights (set based on the amount of data on the edge server or the model's accuracy). The aggregation process is iterative until the model converges.

[0069] Model delivery and update: Global model parameters are encoded as a binary stream and delivered to the edge server via a security protocol (such as TLS). The edge server parses the binary stream and forwards the model to its subordinate terminals. Upon receiving the model, the terminal overwrites the locally stored model file with the global model parameters, completing the update. Simultaneously, the terminal continues local training based on the new model, forming a closed loop.

[0070] The steps for establishing a heterogeneous data collaborative computing model are as follows: Step 1: In the initialization phase, the federated learning server first selects terminals in the computing power network to join the federated learning model training task. After selection, the model is distributed to the terminals to begin training.

[0071] Step 2: After receiving the model, the computing power node first uses local data for training. Considering the heterogeneity of the data, it uploads the distribution of the private data of the computing power terminal to the edge layer node model and merges it with other computing power node models.

[0072] Step 3: After fusing with other computing power node models, input the model into the feature alignment module. After feature alignment, input the model into the edge layer node model, then perform edge layer computing power node fusion, and upload the edge computing power node fusion model to the cloud.

[0073] Step 4: Aggregate the fusion model across a large number of edge computing nodes in the cloud, and then distribute the updated model to each edge computing node. During this process, model consensus is built at the local, group, and global levels to reduce inconsistencies. In each training iteration, the model is optimized using unified adversarial loss, group consensus loss, and global consensus loss. Iterative optimization of the network's intrinsic model effectively alleviates the bias problem of the intrinsic model when dealing with heterogeneous data.

[0074] The efficient joint processing method under the computing power network proposed in this application significantly improves the efficiency and fairness of federated learning under the computing power network through the synergy of soft mask policy gradient and RA-GCN.

[0075] Corresponding to the above method, this application embodiment also provides a high-efficiency joint processing device under a computing power network, as shown in Figure 2. The device includes: a sending unit 210, used to select multiple heterogeneous terminals in the computing power network and send an initial model to each heterogeneous terminal; a training unit 220, used to train the initial model using local data of the corresponding heterogeneous terminals to obtain the trained local model parameters; wherein, the model is adjusted by a training strategy of adjusting the gradient of the soft masking strategy, and the adjustment training strategy includes modeling the computing energy consumption and communication energy consumption of the corresponding heterogeneous terminals; a processing unit 230, used to process the generated local model parameters through a reweighted adversarial graph convolutional network to obtain the model training results of the corresponding heterogeneous terminals; and an aggregation unit 240, used to aggregate the model training results of multiple edge servers and send the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals so that the corresponding heterogeneous terminals update their local model parameters according to the global model.

[0076] The functions of each unit in the high-efficiency joint processing device under a computing network provided in the above embodiments of this application can be implemented through the above-described method steps. Therefore, the specific working process and beneficial effects of each unit in the high-efficiency joint processing device under a computing network provided in the embodiments of this application will not be repeated here.

[0077] This application also provides an electronic device, as shown in FIG3, including a processor 310, a communication interface 320, a memory 330 and a communication bus 340, wherein the processor 310, the communication interface 320 and the memory 330 communicate with each other through the communication bus 340.

[0078] The memory 330 is used to store computer programs; the processor 310 is used to execute the program stored in the memory 330 to implement the implementation steps described in Figure 1.

[0079] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0080] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0081] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0082] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0083] Since the implementation methods and beneficial effects of the various devices in the above embodiments of the electronic device can be achieved by referring to the steps in the embodiment shown in Figure 1, the specific working process and beneficial effects of the electronic device provided in this application embodiment will not be repeated here.

[0084] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform an efficient joint processing method under a computing network as described in any of the above embodiments.

[0085] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the efficient joint processing methods under a computing network described in the above embodiments.

[0086] Those skilled in the art will understand that the embodiments in this application can be provided as methods, systems, or computer program products. Therefore, the embodiments in this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments in this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] This application describes embodiments of methods, apparatus (systems), and computer program products according to embodiments of this application with reference to flowchart illustrations and / or block diagrams. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0089] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0090] Although preferred embodiments have been described in this application, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of this application.

[0091] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of the embodiments of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims in this application and their equivalents, then this application also intends to include these modifications and variations.

Claims

1. A highly efficient joint processing method under a computing power network, characterized in that, The method, applied in a high-efficiency joint processing system including federated learning servers, edge servers, and cloud servers, comprises: the federated learning server selecting multiple heterogeneous terminals in a computing network and distributing an initial model to each heterogeneous terminal; the edge server using local data from the corresponding heterogeneous terminal to train the initial model, obtaining local model parameters generated during training; wherein, the model is adjusted using a training strategy that adjusts the gradient of a soft masking policy, the adjustment strategy including modeling the computational and communication energy consumption of the corresponding heterogeneous terminal; the edge server processing the generated local model parameters through a reweighted adversarial graph convolutional network to obtain the model training results for the corresponding heterogeneous terminal; the cloud server aggregating the model training results from multiple edge servers and distributing the aggregated global model to the edge server, which then forwards it to the corresponding heterogeneous terminal, enabling the corresponding heterogeneous terminal to update its local model parameters according to the global model.

2. The method as described in claim 1, characterized in that, The federated learning server selects multiple heterogeneous terminals from the computing power network through a programmatic algorithm, including: selecting heterogeneous terminals to participate in training from the terminal set using a random selection algorithm or a greedy algorithm based on the terminal's computing power, communication status, or resource availability indicators; and distributing the initial model parameters to the selected heterogeneous terminals in the form of a weight matrix through a network protocol.

3. The method as described in claim 1, characterized in that, Modeling the computational and communication energy consumption of the corresponding heterogeneous terminals includes: Computational energy consumption: calculated based on the number of CPU cycles required to execute one round of training and the computing power of the heterogeneous terminals; wherein, computational energy consumption is directly proportional to the number of CPU cycles and inversely proportional to the computing power; Communication energy consumption: calculated based on the size of the neural network model, the uplink sub-channel bandwidth and the signal-to-noise ratio of the heterogeneous terminals; wherein, communication energy consumption is directly proportional to the model size and inversely proportional to the channel capacity.

4. The method as described in claim 3, characterized in that, The training strategy for adjusting the gradient of the soft masking policy also includes: defining the state space as consisting of training states, available computing resources, optimization metrics, and time delay; defining the action space as consisting of the number of local training iterations on the terminal; and constructing a reward function based on a weighted calculation of model accuracy, energy consumption, and time delay.

5. The method as described in claim 1, characterized in that, Heterogeneous data processing is achieved through a reweighted adversarial graph convolutional network, including: using a weighted network to output sample weights, a classifier to output prediction results, and adversarial training by minimizing the weighted cross-entropy loss and maximizing the regularization term; introducing a feature discriminator to extract feature representations from the local model and outputting feature prediction results through a Softmax function, calculating a unified adversarial loss based on the output probability of the feature discriminator; and constructing a group consensus loss and a global consensus loss, where the group consensus loss is based on KL divergence to calculate the difference between local predictions and group consensus predictions, and the global consensus loss is based on KL divergence to calculate the difference between local predictions and global model predictions.

6. The method as described in claim 5, characterized in that, The formula for calculating the unified adversarial loss is as follows: ,in The output probability of the feature discriminator for sample i on the j-th terminal model is calculated using a logarithmic function and average aggregation.

7. The method as described in claim 1, characterized in that, The cloud server aggregates the model training results from multiple edge servers and distributes the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals. This process includes: aggregating the model parameters uploaded by the edge servers using a weighted average algorithm; encoding the aggregated global model parameters into a binary stream and distributing it to the edge servers via network protocols; the edge servers parsing and forwarding the stream to the corresponding heterogeneous terminals, so that the heterogeneous terminals can update their models by overwriting the locally stored model files after receiving the global model parameters.

8. A high-efficiency joint processing device under a computing power network, characterized in that, The device, applicable to a high-efficiency joint processing system including federated learning servers, edge servers, and cloud servers, comprises: a sending unit for selecting multiple heterogeneous terminals in a computing network and distributing an initial model to each heterogeneous terminal; a training unit for training the initial model using local data from the corresponding heterogeneous terminals to obtain local model parameters; wherein the model is adjusted using a soft masking strategy gradient adjustment training strategy, which includes modeling the computational and communication energy consumption of the corresponding heterogeneous terminals; a processing unit for processing the generated local model parameters using a reweighted adversarial graph convolutional network to obtain the model training results for the corresponding heterogeneous terminals; and an aggregation unit for aggregating the model training results from multiple edge servers and distributing the aggregated global model to the edge servers, which then forward it to the corresponding heterogeneous terminals so that the corresponding heterogeneous terminals update their local model parameters according to the global model.

9. An electronic device, characterized in that, The electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor, when executing the program stored in the memory, implements the method described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images