Edge-cloud heterogeneous model co-modeling efficient federal collaborative learning framework

Abstract

The invention discloses an efficient federal collaborative learning framework for edge-cloud heterogeneous model co-modeling, and belongs to the technical field of federal learning. In order to solve the problems that traditional federated learning is low in one-way knowledge transmission efficiency, privacy protection and model performance are difficult to balance and only isomorphic model parameter aggregation is supported, mutual training among diversified models is realized by combining a cloud edge interactive handshake protocol (CEHP) and an information gain evaluation mechanism (IGVE); and an attention feature fusion mechanism (AFF) is designed to realize feature exchange between heterogeneous models of different targets, and meanwhile, an inverse gradient equalizer is designed to promote dynamic evolution of the models. By means of the multi-target strategy gradient algorithm, the relation between different targets is deeply mined and balanced, and the efficient federal collaborative learning framework is comprehensively and meticulously optimized. According to the method, generalization of the cloud side model and fine granularity of the edge side model are promoted, so that edge-cloud win-win cooperative training is realized.

Description

Technical Field

[0001] This invention belongs to the field of federated learning technology, specifically to an efficient federated collaborative learning framework for the co-evolution of heterogeneous edge-cloud models. It is suitable for scenarios such as medical image analysis, industrial Internet of Things, and autonomous driving. Background Art

[0002] Large language models (such as ChatGPT, Bard, and ChatGLM-6B) have recently made significant progress, revolutionizing the intelligent service landscape across various fields. In fields such as natural language processing, sentiment analysis, creative writing, and conversational agents, these intelligent services have successfully ushered in a new era of human-computer interaction, significantly improving efficiency, productivity, and user satisfaction. Against this backdrop, a growing number of businesses and research institutions are eager to participate in the construction of large-scale models and share in their benefits. The key to achieving this goal lies in designing effective data sharing mechanisms. In response to this, federated learning (FL) has emerged as an innovative distributed machine learning paradigm. Its core advantage lies in enabling collaborative training of multiple models without requiring the sharing of raw data. However, most existing research is based on idealistic assumptions, making it difficult to directly apply in real-world scenarios.

[0003] First, the assumption of homogeneous model training (e.g., the horizontal FL paradigm). This paradigm typically assumes that a single party initiates the model training task and requires all participants to train based on the same model template. At the same time, all potential candidates can participate in the model training activity by bidding and receive corresponding rewards according to the corresponding rules. However, in the actual edge-cloud co-evolution framework, a series of related challenges must be addressed. On the one hand, edge models are limited by their own computing power and small datasets, making it difficult to train and update large and complex models. On the other hand, the cloud side, as the task initiator, faces the following dilemma: (a) publishing a high-precision model as a base model for participants to train, resulting in participants joining only for the high-precision model without sharing the value of their own data; (b) conversely, publishing a low-precision model may hinder the convergence of such distributed model training, ultimately leading to the failure of the collaborative training task.

[0004] Secondly, the tightly coupled model assumption (such as the longitudinal FL paradigm) requires that all participants must conduct online synchronous training. However, because each participant has different data characteristics, the absence of any party will result in an inability to access information related to its characteristics, which in turn leads to degraded or stagnant model performance. This is unacceptable in real-world scenarios, as the side may be temporarily or permanently missing due to network connectivity issues, hardware failures, voluntary withdrawal, or other uncontrollable factors. The existence of these real-world factors makes the tightly coupled model assumption face many challenges in practical applications and difficult to implement smoothly.

[0005] Finally, there's the multi-party objective convergence assumption. This assumption idealistically assumes that all participants share the same training objectives. Unfortunately, in real-world applications, different participants often have diverse and personalized needs, and these differences manifest themselves in many ways. For example, some participants may focus on improving the model's prediction accuracy within a specific domain to provide accurate decision-making for their own business, while others may be more concerned with the model's generalization capabilities, hoping for strong performance across a wide range of scenarios. Furthermore, from a privacy perspective, due to various concerns, such as data privacy and business privacy, each party may not understand each other's specific training objectives. Therefore, the multi-party objective convergence assumption struggles to align with reality in real-world situations and has significant limitations. Summary of the Invention

[0006] To address the challenges of traditional federated learning, such as inefficient one-way knowledge transfer, difficulty balancing privacy protection and model performance, and support for parameter aggregation only for homogeneous models, this paper proposes the Prometheus framework, an efficient federated collaborative learning framework for heterogeneous edge-cloud model co-evolution. This framework can measure the contribution of model features and achieve win-win collaboration among multiple parties.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] An efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution is proposed. The framework integrates two training structures: cloud-to-edge (C2E) and edge-to-cloud (E2C), and three main technical components: 1) a cloud-edge interaction handshake protocol (CEHP); 2) an information gain value evaluation mechanism (IGVE); and 3) an attention feature fusion mechanism (AFF).

[0009] The cloud-to-edge process (C2E) and the edge-to-cloud process (E2C) are two mutually inverse collaborative training processes; the cloud-to-edge process (C2E) is a cloud-side benefit process, while the edge-to-cloud process (E2C) is an edge-side benefit process;

[0010] The cloud side model is represented as Among them, X c represents the training data of the cloud-side model; F c It is the feature extraction map of the cloud-side model; is the cloud-side feature extraction layer of the model, where: W c represents the weight of the fully connected layer, Y c is the prediction target of the model;

[0011] The side model is represented as Among them, E j Indicates the jth th training data of the side model; F e Represents the feature extraction mapping function of the current model; is the entire feature layer of the side model, expressed as: W e , Represent the fully connected layer weight parameters and prediction targets of the current side model respectively;

[0012] The cloud-to-edge process (C2E) includes the cloud-edge interaction handshake protocol (CEHP) to ensure fair interaction. The cloud-to-edge process is specifically initiated by the edge side to request interaction with the cloud side. After the cloud side evaluates the cooperation budget of the edge side, feature extraction is performed to obtain the cloud side feature extraction layer. Without changing the data and goals In this case, the edge and cloud-side features are collaboratively trained; by sharing the advantages of the cloud-side model, the task granularity and accuracy of the edge model are improved;

[0013] The edge-to-cloud process (E2C) includes an information gain evaluation mechanism (IGVE) and feature perturbation technology to ensure fair feature interaction between edge and cloud models. The edge-to-cloud process (E2C) is divided into two cases: "known target" and "unknown target". When the target is a known target, the edge-to-cloud process is essentially the inverse process of the cloud-to-edge process except for the interaction features. The edge-to-cloud process is the edge-side feature after adding noise, and the cloud-to-edge process is the cloud-side feature extraction block after pruning. When the target is an unknown target, feature collaborative training based on the attention feature fusion mechanism (AFF) is used to improve the granularity and accuracy of the edge-side model, while also improving the generalization and diversity of the cloud-side model.

[0014] Furthermore, the Cloud-Edge Interaction Handshake Protocol (CEHP) is a protocol designed specifically for collaborative learning. It is used to ensure privacy protection and model security during cloud-edge collaboration. It dynamically adjusts the collaboration budget and value coefficient to fairly balance the feature learning and benefit exchange relationship between the edge-side model and the cloud-side model, thereby improving the performance and efficiency of the overall system while ensuring privacy. Specifically:

[0015] First, in the cloud-to-edge (C2E) process, the edge side Submit collaborative learning requests to the cloud-side model, which evaluates the requests and selects the ones that meet the requirements. Feature extraction layer Then the cloud side model is based on the budget distance right Perform pruning operations to obtain the revised cloud-side feature extraction block

[0016] Then, the cloud side will Send to the client-side model and request the client to provide its own features Then enter the edge to cloud (E2C) process;

[0017] The edge receives data from the cloud side. Then, the feature block value coefficient θ is evaluated by the information gain evaluation mechanism (IGVE) j ; Side based on θ j Its own characteristics Add Gaussian noise to generate noisy side features And send it to the cloud-side model for collaborative training.

[0018] Furthermore, the information gain evaluation mechanism (IGVE) calculates the information gain between cloud-side features and the prediction target of the edge model. Information gain measures the degree to which the uncertainty of the prediction target is reduced after the introduction of cloud-side features. By quantifying information gain, we determine the extent to which cloud-side features provide valuable information for improving the prediction performance of the edge model. The specific implementation includes initial entropy, cloud-side conditional entropy, and information gain.

[0019] The initial entropy is calculated by calculating the entropy of the target prediction information without using the cloud-side feature layer. This entropy is used to measure the initial uncertainty of the target prediction.

[0020]

[0021] in, represents the training dataset of the edge model, and p(x) represents the probability of the target prediction x;

[0022] The cloud-side conditional entropy is calculated based on the conditional entropy of the edge model target prediction given the cloud-side feature layer, and is expressed as the following formula:

[0023]

[0024] Among them, F represents the cloud-side feature extraction block, Indicates cloud-side feature blocks The probability of Indicates that in a given cloud-side feature block The prediction information entropy of the edge model target under the condition of ; by calculating the conditional entropy, the uncertainty of the edge model's target prediction under the condition of known cloud-side features is quantified;

[0025] The information gain is calculated by comparing the initial information entropy and the conditional entropy of the cloud-side feature layer, and measures the contribution of the cloud-side feature block to the target prediction of the edge model.

[0026]

[0027] in, represents information gain, and θ represents the value coefficient. A larger information gain indicates that the cloud-side feature block contributes more to the target prediction of the edge model, so the value coefficient of the edge model is also higher.

[0028] Furthermore, the cloud-to-edge process (C2E) includes three key components: feature extraction, collaborative training, and fully connected layer fine-tuning; the edge model improves its own performance by leveraging the advantages of the cloud model; Represents the cloud-side feature extraction block;

[0029] In the Prometheus framework, during each round of interaction, the cloud-side model provides the feature extraction layer for the edge-side model for collaborative training. After evaluating the value of the cloud-side feature blocks, the edge model extracts features based on the collaborative budget request and performs general training on the cloud-side feature blocks. This round of value evaluation returns the remaining collaborative budget to prepare for the next round of training. During each update, the cloud-side and edge models learn from each other to improve performance.

[0030]

[0031] Among them, P represents the pruning operation, Δ value Indicates the degree of pruning applied to the feature extraction block of the cloud-side model;

[0032] The collaborative training process described above: During the cloud-to-edge (C2E) process, the cloud-side feature extraction block is combined with the edge-side features and trained together with the original edge-side fully connected layer. During the cloud-to-edge (C2E) collaborative training, the parameters of the cloud-side feature extraction block are fixed, and only the features of the edge model itself are updated. Throughout the entire process, the edge model absorbs the knowledge shared by the cloud side to achieve improved task granularity.

[0033] Define a collaborative training convolutional network as:

[0034] in, and They represent the cloud-side feature blocks and the edge-side feature blocks respectively, and ⊙ represents the feature collaborative training process.

[0035] Furthermore, the training focus of the edge-to-cloud (E2C) process shifts from the edge to the cloud, with features provided by the edge and data provided by the cloud. To ensure fair interaction between the cloud and edge, the cloud-side feature extraction module performs pruning in the cloud-to-edge (C2E) process, while edge features are noise-added in the E2C process. The specific definitions are as follows:

[0036] In the feature extraction module, the edge side will evaluate the cloud side feature extraction block through information gain The value coefficient θ j ; This value coefficient is used to determine the appropriate noise addition to the side model features to prevent feature leakage;

[0037]

[0038] in, is the side feature after noise processing, The mean is 0 and the standard deviation is Gaussian noise.

[0039] Furthermore, when the target is unknown, feature collaborative training based on the Attention Feature Fusion (AFF) mechanism is used to improve the granularity and accuracy of the edge model, while also enhancing the generalization and diversity of the cloud model. This is achieved through four parts: an edge-cloud interconnection verification mechanism, a cooperative budget update design, an unknown target collaborative training scenario, and reinforcement learning optimization.

[0040] The edge-cloud interaction verification mechanism includes two key components: the cloud-edge interaction handshake protocol (CEHP) and the information gain evaluation mechanism (IGVE).

[0041] Furthermore, the collaborative budget update design specifically performs collaborative training between the cloud side and the edge side through a collaborative budget update method while protecting the target privacy;

[0042] Assuming there are n edge models, in the initial iteration of training, the cloud side will randomly initialize the collaboration budget of each edge model To meet the conditions Subsequently, based on the cloud-edge interaction handshake protocol, edge-cloud collaborative interaction is carried out, and each handshake process represents an interaction; collaborative training between models is an iterative process that requires multiple rounds of interaction, so an edge-cloud coordination point is set up to achieve fairness and privacy protection in multiple rounds of collaborative interaction training; in the proposed edge-cloud collaborative training round, each edge feature interacts with its corresponding cloud-side feature extraction block to form a pair of edge-cloud feature pairs; the value evaluation coefficient θ of each pair of edge-cloud features j Represents the degree of correlation between the two;

[0043] The feedback mechanism introduced includes the reverse gradient balancing mechanism and the collaborative budget recovery mechanism;

[0044] The reverse gradient balancing mechanism is based on the value evaluation coefficient θ of each edge model j Redistribute the collaboration budget; there are m edge models, and their corresponding value evaluation coefficients are θ j ,j=1,…,m; the cooperation budget will be based on these value assessment coefficients After training is completed, if there are m edge models waiting for collaboration budget reallocation, the cloud side will reallocate them based on the reallocation coefficient. Redistribute the collected collaboration budget in reverse order;

[0045] The collaborative budget recovery mechanism is based on θ j The gradient value of the collaboration budget is recovered; if the θ of a certain edge model j The gradient change is close to zero, indicating that its performance improvement has reached its limit. At this time, the cloud side can partially reclaim the collaboration budget of the edge model and reallocate it to other edge models waiting for training; the edge model will retain part of the collaboration budget to continue collaborative training with other larger models.

[0046] Furthermore, the unknown target collaborative training scenario is specifically:

[0047] Assume there are n edge features and one cloud side target; use θ1,θ2,…,θ n Represents the information gain coefficient of each side model feature, using a1, a2, ..., a n Represents its corresponding attention weight; the attention weight a of each side model feature j j The calculation formula is as follows: Given the feature maps of the cloud side and the edge side Introducing attention weights to achieve attention-based feature fusion between cloud-side models and edge models;

[0048] in, Represents the fused features, C represents the number of channels, H represents the number of samples, and W represents the dimension of the features; Represents the attention weight generated by MS-CAM. Introducing a scalar value a j As the updated attention weight, where a j ∈[0,1] and 1-a j This constraint is also satisfied; the network can and Soft selection or weighted averaging between them.

[0049] Furthermore, the reinforcement learning optimization is specifically:

[0050] Edge-cloud collaborative training is expressed as a Markov decision process (MDP), using the policy gradient π(a|λ c ,λ e ;ε,θ) to optimize the collaborative strategy;

[0051] Among them, the policy gradient consists of two parts: the first part is: π c (i|λ c ; ε e ) According to the feature map λ in the cloud-side feature unit c Select the feature extraction block for co-training; the second part is: π e (j|λ e ; θ) is used to select appropriate end-side features for collaborative training; these two parts are integrated into a loss function for optimization:

[0052] Part I, Strategy π c Defined by the cloud-side model prediction accuracy: Where i = 1,…,N, is the i-th score of the cloud-side collaborative training effect; the return rate during the holding period t is in is the rate of increase of prediction accuracy; given the initial score The cumulative return of a trajectory τ is: The optimization goal of this part is to maximize the logarithmic cumulative benefit of all trajectories τ:

[0053] Part II, Strategy π e Defined by the edge-side model prediction accuracy: Where, j = 1,…,M, is the jth score of the edge-cloud collaborative training effect; similar to the cloud side, the return rate for holding period t is Given an initial score The cumulative gain of a trajectory is: Therefore, the optimization objective of the side model is: That is, the overall objective function is and The weighted sum of: Among them, ζ is a hyperparameter used to control the different learning rates of each part; these two parts are Optimize at the same time.

[0054] An efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution with applications in medical image analysis, industrial Internet of Things, or autonomous driving.

[0055] Compared with the prior art, the present invention has the following advantages:

[0056] In terms of quantifying the characteristic value of participants, the present invention first uses information gain to quantify the characteristic value of each side participant.

[0057] Based on this, we further introduce a value coefficient, which is used to further infer the feature contribution of the current side model throughout the training process. Based on the inferred value coefficient, an inverse gradient equalizer (InverseGrad Equalizer) is established to redistribute the cooperation budgets of participants. At its core, the side model submits its self-determined value coefficient to the cloud side. The cloud side, as the allocator, redistributes the cooperation budget based on this value coefficient. This provides a fair and just interaction environment for all parties in federated learning, ensuring that resource allocation matches the contributions of each party and avoiding collaborative imbalances caused by uneven resource allocation.

[0058] To address the widespread problem of incompatible goals in federated learning, this paper innovatively proposes a new collaborative configuration scheme for unknown-goal models. Through the design of a heterogeneous training model architecture and strategy, it enables participants with different goal orientations to achieve effective collaboration within the same federated learning framework. This allows for collaborative training while preserving privacy, breaking down the barriers to collaboration that arise from divergent goals in traditional federated learning.

[0059] Finally, based on the principle of diminishing marginal returns, this paper meticulously designs a novel reward-penalty mechanism. This mechanism fully accounts for the fact that as participants gradually increase their resources or behaviors, the incremental benefits gained per unit of investment tend to decrease, thereby more precisely constructing a system of reward and penalty rules. Furthermore, by leveraging a multi-objective policy gradient algorithm, we deeply explore and balance the relationships between different objectives, enabling comprehensive and meticulous optimization of Prometheus. This approach promotes the generalization of cloud-side models and the fine-grainedness of edge-side models, thereby achieving a win-win collaborative training between edge and cloud. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0061] Figure 1 This is a schematic diagram of the overall structure of Prometheus;

[0062] Figure 2 This is the Prometheus model architecture diagram;

[0063] Figure 3 Schematic diagram of attention fusion mechanism

[0064] Figure 4 This is a comparison chart of the model convergence effect;

[0065] Figure 5 This is a schematic diagram comparing heterogeneity challenges;

[0066] Figure 6 It is the experimental picture of the target unknown scene. DETAILED DESCRIPTION

[0067] To gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description thereof. However, the present invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a comprehensive understanding of the present disclosure.

[0068] An efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution. This framework integrates two training structures:

[0069] Cloud-to-edge process (C2E) and edge-to-cloud process (E2C), as well as three main technical components: 1) Cloud-Edge Interaction Handshake Protocol (CEHP); 2) Information Gain Value Evaluation Mechanism (IGVE); 3) Attention Feature Fusion Mechanism (AFF);

[0070] The cloud-to-edge process (C2E) and the edge-to-cloud process (E2C) are two mutually inverse collaborative training processes; the cloud-to-edge process (C2E) is a cloud-side benefit process, while the edge-to-cloud process (E2C) is an edge-side benefit process;

[0071] The cloud side model is represented as Among them, X c represents the training data of the cloud-side model; F C It is the feature extraction map of the cloud-side model; is the cloud-side feature extraction layer of the model, where: W c represents the weight of the fully connected layer, Y c is the prediction target of the model;

[0072] The side model is represented as Among them, E j Indicates the jth th training data of the side model; F e Represents the feature extraction mapping function of the current model; is the entire feature layer of the side model, expressed as: W e , Represent the fully connected layer weight parameters and prediction targets of the current side model respectively;

[0073] The cloud-to-edge process (C2E) includes the cloud-edge interaction handshake protocol (CEHP) to ensure fair interaction. The cloud-to-edge process is specifically initiated by the edge side to request interaction with the cloud side. After the cloud side evaluates the cooperation budget of the edge side, feature extraction is performed to obtain the cloud side feature extraction layer. Without changing the data and goals In this case, the edge and cloud-side features are collaboratively trained; by sharing the advantages of the cloud-side model, the task granularity and accuracy of the edge model are improved;

[0074] We also propose an intelligent cooperation budget allocation mechanism, where the cloud-side model dynamically allocates cooperation budgets based on the budget requests of each edge before each round of training. Specifically, in the tth round of cloud-to-edge process (C2E), any edge client E that is allocated a cooperation budget ε j Submit collaborative interaction requests to the cloud side (i.e. cooperation budget ε). A specific feature extraction layer is provided to the edge. By sharing the advantages of the large cloud-side model, the fine-grained nature of the edge model is enhanced. The goal is to reduce training costs and achieve faster and more efficient convergence, thereby sharing the benefits of the large cloud-side model. This process represents the profit-making process of the edge model.

[0075]

[0076] Among them, Δ i is the collaborative budget threshold for feature extraction on the cloud side; represents the cooperative budget allocation value; σ pr Represents the pruning operation in the cloud-side feature extraction process; Represents the model validation process; represents the feature co-training operation; ξ is an adjustable bound during model evaluation.

[0077] The edge-to-cloud process (E2C) includes an information gain evaluation mechanism (IGVE) and feature perturbation technology to ensure fair feature interaction between edge and cloud models. The edge-to-cloud process (E2C) is divided into two cases: "known target" and "unknown target". When the target is a known target, the edge-to-cloud process is essentially the inverse process of the cloud-to-edge process except for the interaction features. The edge-to-cloud process is the edge-side feature after adding noise, and the cloud-to-edge process is the cloud-side feature extraction block after pruning. When the target is an unknown target, feature collaborative training based on the attention feature fusion mechanism (AFF) is used to improve the granularity and accuracy of the edge-side model, while also improving the generalization and diversity of the cloud-side model.

[0078] During each training round, when the cloud-side feature extraction module provides extracted features to a specific edge-side model, the edge-side model then quantifies the information gain of the received features, deriving a value coefficient that accurately reflects the value of these features. Based on this information gain, the edge-side model injects corresponding Gaussian noise into its entire feature layer and transmits it back to the cloud. This provides the cloud with fine-grained knowledge, enhancing the cloud-side model's generalization capabilities and performance on specific downstream tasks, thereby achieving "model self-learning." This process represents the profit-making process of the cloud-side model.

[0079]

[0080] in, represents a new data set; h c (D) shows the performance of the model on the dataset. G c This measures the average performance of the model on new data, i.e., its generalization ability. A lower generalization error indicates that the model is more suitable for unseen data and has better generalization ability.

[0081] Furthermore, the Cloud-Edge Interaction Handshake Protocol (CEHP) is a protocol designed specifically for collaborative learning. It is used to ensure privacy protection and model security during cloud-edge collaboration. It dynamically adjusts the collaboration budget and value coefficient to fairly balance the feature learning and benefit exchange relationship between the edge-side model and the cloud-side model, thereby improving the performance and efficiency of the overall system while ensuring privacy. Specifically:

[0082] First, in the cloud-to-edge (C2E) process, the edge side Submit collaborative learning requests to the cloud-side model, which evaluates the requests and selects the ones that meet the requirements. Feature extraction layer Then the cloud side model is based on the budget distance right Perform pruning operations to obtain the revised cloud-side feature extraction block

[0083] Then, the cloud side will Send to the client-side model and request the client to provide its own features Then enter the edge to cloud (E2C) process;

[0084] The edge receives data from the cloud side. Then, the feature block value coefficient θ is evaluated by the information gain evaluation mechanism (IGVE) j ; Side based on θ j Its own characteristics Add Gaussian noise to generate noisy side features And send it to the cloud-side model for collaborative training.

[0085] Furthermore, the information gain evaluation mechanism (IGVE) calculates the information gain between cloud-side features and the prediction target of the edge model. Information gain measures the degree to which the uncertainty of the prediction target is reduced after the introduction of cloud-side features. By quantifying information gain, we determine the extent to which cloud-side features provide valuable information for improving the prediction performance of the edge model. The specific implementation includes initial entropy, cloud-side conditional entropy, and information gain.

[0086] The initial entropy is calculated by calculating the entropy of the target prediction information without using the cloud-side feature layer. This entropy is used to measure the initial uncertainty of the target prediction.

[0087]

[0088] in, represents the training dataset of the edge model, and p(x) represents the probability of the target prediction x;

[0089] The cloud-side conditional entropy is calculated based on the conditional entropy of the edge model target prediction given the cloud-side feature layer, and is expressed as the following formula:

[0090]

[0091] Among them, F represents the cloud-side feature extraction block, Indicates cloud-side feature blocks The probability of Indicates that in a given cloud-side feature block The prediction information entropy of the edge model target under the condition of ; by calculating the conditional entropy, the uncertainty of the edge model's target prediction under the condition of known cloud-side features is quantified;

[0092] The information gain is calculated by comparing the initial information entropy and the conditional entropy of the cloud-side feature layer, and measures the contribution of the cloud-side feature block to the target prediction of the edge model.

[0093]

[0094] in, represents information gain, and θ represents the value coefficient. A larger information gain indicates that the cloud-side feature block contributes more to the target prediction of the edge model, so the value coefficient of the edge model is also higher.

[0095] Furthermore, the cloud-to-edge process (C2E) includes three key components: feature extraction, collaborative training, and fully connected layer fine-tuning; the edge model improves its own performance by leveraging the advantages of the cloud model; Represents the cloud-side feature extraction block;

[0096] In the Prometheus framework, during each round of interaction, the cloud-side model provides the feature extraction layer for the edge-side model for collaborative training. After evaluating the value of the cloud-side feature blocks, the edge model extracts features based on the collaborative budget request and performs general training on the cloud-side feature blocks. This round of value evaluation returns the remaining collaborative budget to prepare for the next round of training. During each update, the cloud-side and edge models learn from each other to improve performance.

[0097]

[0098] Among them, P represents the pruning operation, Δ value Indicates the degree of pruning applied to the feature extraction block of the cloud-side model;

[0099] The collaborative training process described above: During the cloud-to-edge (C2E) process, the cloud-side feature extraction block is combined with the edge-side features and trained together with the original edge-side fully connected layer. During the cloud-to-edge (C2E) collaborative training, the parameters of the cloud-side feature extraction block are fixed, and only the features of the edge model itself are updated. Throughout the entire process, the edge model absorbs the knowledge shared by the cloud side to achieve improved task granularity.

[0100] Define a collaborative training convolutional network as:

[0101] in, and They represent the cloud-side feature blocks and the edge-side feature blocks respectively, and ⊙ represents the feature collaborative training process.

[0102] Furthermore, the training focus of the edge-to-cloud (E2C) process shifts from the edge to the cloud, with features provided by the edge and data provided by the cloud. To ensure fair interaction between the cloud and edge, the cloud-side feature extraction module performs pruning in the cloud-to-edge (C2E) process, while edge features are noise-added in the E2C process. The specific definitions are as follows:

[0103] In the feature extraction module, the edge side will evaluate the cloud side feature extraction block through information gain The value coefficient θ j ; This value coefficient is used to determine the appropriate noise addition to the side model features to prevent feature leakage;

[0104]

[0105] in, is the side feature after noise processing, The mean is 0 and the standard deviation is Gaussian noise.

[0106] The following algorithm shows the process of collaborative training between the edge and the cloud:

[0107]

[0108] This collaborative training model fully integrates cloud-side and edge-side resources and knowledge, improving the efficiency and effectiveness of model training. By introducing noise terms and setting up a conditional judgment mechanism, the algorithm can ensure the model's robustness and adaptability to a certain extent, effectively avoiding overfitting and the negative impact of poor edge-side models on the overall training process.

[0109] Furthermore, when the target is unknown, feature collaborative training based on the Attention Feature Fusion (AFF) mechanism is used to improve the granularity and accuracy of the edge model, while also enhancing the generalization and diversity of the cloud model. This is achieved through four parts: an edge-cloud interconnection verification mechanism, a cooperative budget update design, an unknown target collaborative training scenario, and reinforcement learning optimization.

[0110] The edge-cloud interaction verification mechanism includes two key components: the cloud-edge interaction handshake protocol (CEHP) and the information gain evaluation mechanism (IGVE).

[0111] Furthermore, the collaborative budget update design specifically performs collaborative training between the cloud side and the edge side through a collaborative budget update method while protecting the target privacy;

[0112] Assuming there are n edge models, in the initial iteration of training, the cloud side will randomly initialize the collaboration budget of each edge model To meet the conditions Subsequently, based on the cloud-edge interaction handshake protocol, edge-cloud collaborative interaction is carried out, and each handshake process represents an interaction; collaborative training between models is an iterative process that requires multiple rounds of interaction, so an edge-cloud coordination point is set up to achieve fairness and privacy protection in multiple rounds of collaborative interaction training; in the proposed edge-cloud collaborative training round, each edge feature interacts with its corresponding cloud-side feature extraction block to form a pair of edge-cloud feature pairs; the value evaluation coefficient θ of each pair of edge-cloud features j Represents the degree of correlation between the two;

[0113] The feedback mechanism introduced includes the reverse gradient balancing mechanism and the collaborative budget recovery mechanism;

[0114] The reverse gradient balancing mechanism is based on the value evaluation coefficient θ of each edge model j Redistribute the collaboration budget; there are m edge models, and their corresponding value evaluation coefficients are θ j ,j=1,…,m; the cooperation budget will be based on these value assessment coefficients After training is completed, if there are m edge models waiting for collaboration budget reallocation, the cloud side will reallocate them based on the reallocation coefficient. Redistribute the collected collaboration budget in reverse order;

[0115] The collaborative budget recovery mechanism is based on θ j The gradient value of the collaboration budget is recovered; if the θ of a certain edge model j The gradient change is close to zero, indicating that its performance improvement has reached its limit. At this time, the cloud side can partially reclaim the collaboration budget of the edge model and reallocate it to other edge models waiting for training; the edge model will retain part of the collaboration budget to continue collaborative training with other larger models.

[0116] Furthermore, the unknown target collaborative training scenario is specifically:

[0117] Assume there are n edge features and one cloud side target; use θ1,θ2,…,θ n Represents the information gain coefficient of each side model feature, using a1, a2, ..., a n Represents its corresponding attention weight; the attention weight a of each side model feature j j The calculation formula is as follows: Given the feature maps of the cloud side and the edge side Introducing attention weights to achieve attention-based feature fusion between cloud-side models and edge models;

[0118] in, Represents the fused features, C represents the number of channels, H represents the number of samples, and W represents the dimension of the features; Represents the attention weight generated by MS-CAM. Introducing a scalar value a j As the updated attention weight, where a j ∈[0,1] and 1-a j This constraint is also satisfied; the network can and Soft selection or weighted averaging between them.

[0119] The specific algorithm is as follows:

[0120]

[0121] Furthermore, the reinforcement learning optimization is specifically:

[0122] Edge-cloud collaborative training is expressed as a Markov decision process (MDP), using the policy gradient π(a|λ c ,λ e ;ε,θ) to optimize the collaborative strategy;

[0123] Among them, the policy gradient consists of two parts: the first part is: π c (i|λ c ; ε e ) According to the feature map λ in the cloud-side feature unit c Select the feature extraction block for co-training; the second part is: π e (j|λ e ; θ) is used to select appropriate end-side features for collaborative training; these two parts are integrated into a loss function for optimization:

[0124] Part I, Strategy π c Defined by the cloud-side model prediction accuracy: Where i = 1,…,N, is the i-th score of the cloud-side collaborative training effect; the return rate during the holding period t is in is the rate of increase of prediction accuracy; given the initial score The cumulative return of a trajectory τ is: The optimization goal of this part is to maximize the logarithmic cumulative benefit of all trajectories τ:

[0125] Part II, Strategy π e Defined by the edge-side model prediction accuracy: Where, j = 1,…,M, is the jth score of the edge-cloud collaborative training effect; similar to the cloud side, the return rate for holding period t is Given an initial score The cumulative gain of a trajectory is: Therefore, the optimization objective of the side model is: That is, the overall objective function is and The weighted sum of: Among them, ζ is a hyperparameter used to control the different learning rates of each part; these two parts are Optimize at the same time.

[0126] Assessment Test

[0127] The MNIST, CIFAR-10, and CIFAR-100 datasets were used to evaluate the capabilities of Prometheus in this application.

[0128] The MNIST dataset contains handwritten digits from 0 to 9, divided into 10 categories, with 60,000 images for training and 10,000 for testing. Each category of digits contains 6,000 training images and 1,000 test images, and all images are 28×28 pixels in size.

[0129] The CIFAR-10 and CIFAR-100 datasets contain images from 10 and 100 categories, respectively. The CIFAR-10 dataset consists of 50,000 training images and 10,000 test images, with 5,000 training images and 1,000 test images per class. The CIFAR-100 dataset has the same total number of images as CIFAR-10, but is expanded to 100 categories, with 500 training images and 100 test images per class. Images in the CIFAR-10 / 100 datasets are represented in color using three channels, with each image measuring 32×32 pixels. These datasets serve as benchmarks for Prometheus experiments.

[0130] Baseline models: Two basic methods in federated learning (FL) are used as baseline models: FedAvg and FedProx. FedAvg is a benchmark algorithm for federated learning. It works by aggregating the parameters of locally trained models from clients through weighted averaging to build a unified global model. All models, including those deployed in the cloud and those customized for individual clients, maintain architectural consistency. FedProx introduces a key innovation based on FedAvg to address the problem of data heterogeneity between devices: a proximal regularization term is integrated into the loss function. This term constrains local model updates to keep them consistent with the parameters of the global model, thereby promoting convergence to the aggregated mean of the locally trained model and mitigating the adverse effects of non-independent and identically distributed (non-IID) data distributions. Both models follow the same architectural design and are uniformly deployed in the cloud and client environments to ensure fair comparison.

[0131] Federated Task Setup: The experimental workflow is executed in a simulated cross-silo federated learning (FL) environment supervised by a central coordinator that directs the operations of five participating clients (K=5). In the model heterogeneity experiments, a fixed number of communication rounds is specified. For the MNIST and CIFAR-100 datasets, the rounds are set to 100, while for the CIFAR-10 dataset, the rounds are set to 200. In the target heterogeneity experiments, the total number of communication rounds is set to 250. In all scenarios, the number of local iterations (or epochs) is fixed to 10. To ensure fair distribution of learning data among participants, the dataset is divided into five equal parts for IID and non-IID environments. This distribution ensures that each participant receives an equal share of training and test data, with each participant having one-fifth of the training data and one-fifth of the test data for private validation. For example, in the CIFAR-10 dataset, when 50,000 training images and 10,000 test images are allocated under the IID condition, each participant will receive 10,000 training images and 2,000 test images for private training and validation, and there is no overlap in the initial data allocation. The effectiveness of the global model is evaluated using a comprehensive test suite containing 10,000 test images from the CIFAR-10 dataset. In the IID (independent and identically distributed) scenario, following the standard federated learning setting, participants are given a random subset of the data to ensure a uniform distribution across all classes, denoted as P i (x,y)=P joint(x,y). In contrast, in the Non-IID framework, the dataset is divided into Kp different segments, each of size n / Kp, where n and K represent the total number of clients and the total size of the dataset, respectively (in this case K = 5, n = 50,000), and each client is assigned p shares of data. The experimental design considers three gradually increasing levels of Non-IID complexity, represented by different p values. For the MNIST and CIFAR-10 datasets, p takes the values of {6, 4, 2}, while for the CIFAR-100 dataset, p is set to {60, 40, 20}. This means that the maximum number of classes assigned to each client is "p". These experimental conditions are labeled Non-IID (1, 2, 3), where the Non-IID (3) scenario represents the most challenging level with no class overlap between clients.

[0132] Network structure setting: The performance of four different modeling architectures was evaluated in the experiment: multi-layer perceptron (MLP), LeNet5, a convolutional neural network (CNN1) containing three 3×3 convolutional layers (initially with six filters, gradually increasing to 12, and finally 16), each followed by a 2×2 maximum pooling layer and a rectified linear unit (ReLU) activation function, and another CNN variant (CNN2) with three 128-filter 3×3 convolutional layers, each followed by a 2×2 maximum pooling layer and a ReLU activation and a fully connected layer. The MNIST dataset is paired with the MLP and LeNet5 models, while CNN1 and CNN2 are used for the CIFAR-10 / 100 dataset. The optimization strategy for all models uses the stochastic gradient descent (SGD) algorithm with a momentum parameter of 0.9 and a weight decay of 5×10 -4 , with a batch size of 128. Experiments were conducted on an NVIDIA GeForce RTX 2080Ti GPU, with a learning rate of 0.001 for the CNN1 and CNN2 models and a learning rate of 0.01 for the MLP and LeNet5 models. The learning rate was decayed by a factor of 0.1 every 10 epochs. The experiment ran for a total of 100 epochs, with early stopping for 5 epochs when there was no improvement on the validation set.

[0133] Comparative experiment

[0134] The Prometheus of the present invention is compared with two classic federated learning (FL) benchmark models FedAvg and FedProx to evaluate the effectiveness of Prometheus.

[0135] In this experiment, comparisons were made in independent and identically distributed (IID) and non-independent and identically distributed (Non-IID) scenarios, focusing on model performance, convergence speed, and overall stability under different data distributions.

[0136] The experimental environment consists of a central coordinator and five clients, using MNIST, CIFAR-10, and CIFAR-100 as datasets. The experiments are conducted under three levels of Non-IID settings (1, 2, and 3). Table 1 shows the Top-1 accuracy of the global model trained on different datasets using FedAvg, FedProx, and Prometheus. The results show that Prometheus outperforms FedAvg and FedProx on most datasets 4E0A, and its advantage is particularly significant in the non-IID scenario. In the non-IID (3) setting, the accuracy of FedAvg and FedProx drops severely, while Prometheus mitigates the negative impact of data heterogeneity through its edge-cloud collaborative learning mechanism and maintains a high global model accuracy.

[0137] Table 1 Top-1 accuracy of global models under typical federated learning (FL) settings

[0138]

[0139] The impact of different federated learning frameworks is further analyzed by comparing the training convergence curves of the three methods. Figure 4 As shown, Prometheus demonstrates faster and more stable training convergence under all data distributions. In contrast, FedAvg converges slowly and suffers from training oscillations in the non-IID setting. While FedProx improves stability, its accuracy remains lower than Prometheus. Experimental data shows that in the non-IID setting, Prometheus achieves 3.2% higher accuracy than FedAvg and 4.1% higher accuracy than FedProx, demonstrating its effectiveness in heterogeneous data environments. Furthermore, Prometheus achieves lower training variance, demonstrating improved robustness and generalization capabilities in the face of imbalanced data.

[0140] Model validation experiments

[0141] Verify the core mechanism of Prometheus and improve the model performance of the edge model and the generalization ability of the cloud model through edge-cloud collaboration. Select the CIFAR-100 dataset as the training dataset and use ResNet-18

[81] CNN1 and CNN2 serve as the cloud-side models, while CNN2 and CNN3 serve as the edge-side models. First, baseline training is performed for each model, and its performance on the training data is recorded. Subsequently, joint training is performed using the Prometheus framework to ensure collaborative learning between the edge and cloud-side models. The performance gains from the two training methods are recorded and compared, and the following performance metrics are evaluated: accuracy, FLOPs (floating-point operations per second), and F1-score. For each model, the baseline training results are compared with those obtained using Prometheus. As shown in Table 2, Prometheus significantly improves learning efficiency. For example, in the ResNet-18 (cloud-side model) experiment, the accuracy increases from 89.67% to 93.46%, while the accuracy of CNN1 and CNN2 (edge-side models) increases by 2.77% and 0.64%, respectively. Furthermore, Prometheus improves F1-score while optimizing computational cost, demonstrating its effectiveness in stabilizing predictions and enhancing generalization.

[0142] Table 2 Model improvement effect

[0143]

[0144] A collaborative training approach is adopted to address the challenge of model heterogeneity, enabling each client to customize the model according to its specific needs.

[0145] Experimental setup: Each of the five clients is assigned a unique model. Specifically, client 1 uses the multi-layer perceptron (MLP) model, client 2 uses the LENET5 architecture, clients 3 and 4 use the CNN1 architecture, and the last client 5 uses the CNN2 architecture; the overall model uses RESNET-18. The model is trained using the CIFAR-10 dataset, ensuring that the data is independent and identically distributed (IID). The experimental results are shown in the figure below. Figure 5 As shown, we evaluate the performance of a personalized model trained using Prometheus (represented by the solid curve) and compare it to the highest accuracy achieved by training an independent model using an independent validation set (represented by the dashed line). We show the first 100 epochs of the training process.

[0146] like Figure 6As shown, it can be observed that the model architecture plays a key role in the difference in accuracy. The models based on convolutional neural networks (CNN) (clients 3, 4, and 5) achieved higher accuracy due to their excellent feature extraction capabilities, while the shallower models of multi-layer perceptron (client 1) and LeNet5 (client 2) had difficulties in processing complex image representations, resulting in lower performance. In addition, federated collaboration is more beneficial to deeper models, which can utilize diverse feature learning from different clients. The accuracy fluctuations of PRM-client1 and PRM-client5 indicate that some models are more sensitive to federated updates and changes in local datasets. In summary, the experimental results confirm that the Prometheus of the present invention effectively enhances personalized training, but the degree of improvement depends on the complexity of the model and its generalization ability in a federated environment.

[0147] In the adaptability experiment for heterogeneous task targets, Prometheus was tested in the scenario of assigning different classification tasks to clients.

[0148] Any matters not described in detail in this specification are prior art known to those skilled in the art. Although the above description of the present invention is based on specific embodiments to facilitate understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the specific embodiments. As long as various modifications are within the spirit and scope of the present invention as defined and determined by the appended claims, such modifications will be obvious to those skilled in the art, and all inventions and creations utilizing the concepts of the present invention are protected.

Claims

1. An efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution, characterized by: The proposed efficient federated collaborative learning framework is an edge-cloud collaborative training framework that integrates two training structures: cloud-to-edge and edge-to-cloud, and three main technical components: 1) a cloud-edge interaction handshake protocol; 2) an information gain value evaluation mechanism; and 3) an attention feature fusion mechanism. The cloud-to-edge process and the edge-to-cloud process are two mutually inverse collaborative training processes; the cloud-to-edge process is a process that benefits the cloud side, while the edge-to-cloud process is a process that benefits the edge side; The cloud side model is represented as Among them, X c represents the training data of the cloud-side model; F c It is the feature extraction map of the cloud-side model; is the cloud-side feature extraction layer of the model, where: W c represents the weight of the fully connected layer, Y c is the prediction target of the model; The side model is represented as Among them, E j Indicates the jth th training data of the side model; F e Represents the feature extraction mapping function of the current model; is the entire feature layer of the side model, expressed as: W e , Represent the fully connected layer weight parameters and prediction targets of the current side model respectively; The cloud-to-edge process includes a cloud-edge interaction handshake protocol to ensure fair interaction. The cloud-to-edge process is specifically initiated by the edge side to request interaction with the cloud side. After the cloud side evaluates the edge side's cooperation budget, feature extraction is performed to obtain the cloud side feature extraction layer. Without changing the data and goals In this case, the edge and cloud-side features are collaboratively trained; by sharing the advantages of the cloud-side model, the task granularity and accuracy of the edge model are improved; The edge-to-cloud process includes an information gain evaluation mechanism and feature perturbation technology to ensure fair feature interaction between the edge and cloud models. The edge-to-cloud process is divided into two cases: "known target" and "unknown target". When the target is a known target, the edge-to-cloud process is essentially the inverse of the cloud-to-edge process, except for the interaction features. The edge-to-cloud process is the edge-side feature after adding noise, and the cloud-to-edge process is the cloud-side feature extraction block after pruning. When the target is an unknown target, feature collaborative training based on the attention feature fusion mechanism is used to improve the granularity and accuracy of the edge-side model, while also improving the generalization and diversity of the cloud-side model.

2. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 1 is characterized by: The cloud-edge interaction handshake protocol is a protocol designed specifically for collaborative learning, used to ensure privacy protection and model security during cloud-edge collaboration. It dynamically adjusts the collaboration budget and value coefficient to fairly balance the feature learning and benefit exchange relationship between the edge and cloud models, while ensuring privacy while improving the performance and efficiency of the overall system. Specifically: First, in the cloud-to-edge process, the edge side Submit collaborative learning requests to the cloud-side model, which evaluates the requests and selects the ones that meet the requirements. Feature extraction layer Then the cloud side model is based on the budget distance right Perform pruning operations to obtain the revised cloud-side feature extraction block Then, the cloud side will Send to the client-side model and request the client to provide its own features Then enter the edge-to-cloud process; The edge receives data from the cloud side. Then, the feature block value coefficient θ is evaluated through the information gain evaluation mechanism j ; Side based on θ j Its own characteristics Add Gaussian noise to generate noisy side features And send it to the cloud-side model for collaborative training.

3. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 1 is characterized by: The information gain evaluation mechanism calculates the information gain between cloud-side features and the prediction target of the edge model. Information gain measures the degree to which the uncertainty of the prediction target is reduced after the introduction of cloud-side features. By quantifying information gain, we determine the extent to which cloud-side features provide valuable information for improving the prediction performance of the edge model. The specific implementation includes initial entropy, cloud-side conditional entropy, and information gain. The initial entropy is calculated by calculating the entropy of the target prediction information when the edge model does not use the cloud-side feature layer. This entropy is used to measure the initial uncertainty of the target prediction. in, represents the training dataset of the edge model, and p(x) represents the probability of the target prediction x; The cloud-side conditional entropy is calculated based on the conditional entropy of the edge model target prediction given the cloud-side feature layer, and is expressed as the following formula: Among them, F represents the cloud-side feature extraction block, Indicates cloud-side feature blocks The probability of Indicates that in a given cloud-side feature block The prediction information entropy of the edge model target under the condition of ; by calculating the conditional entropy, the uncertainty of the edge model's target prediction under the condition of known cloud-side features is quantified; The information gain is calculated by comparing the initial information entropy and the conditional entropy of the cloud-side feature layer, and measures the contribution of the cloud-side feature block to the target prediction of the edge model. in, represents information gain, and θ represents the value coefficient. A larger information gain indicates that the cloud-side feature block contributes more to the target prediction of the edge model, so the value coefficient of the edge model is also higher.

4. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 1 is characterized by: The cloud-to-edge process includes feature extraction, collaborative training, and fine-tuning of fully connected layers. The edge model leverages the strengths of the cloud model to improve its performance. The feature extraction described: the convolutional structure naturally contains structural layers; using Represents the cloud-side feature extraction block; In the framework, during each round of interaction, the cloud-side model provides the feature extraction layer for the edge-side model to be used for collaborative training. After evaluating the value of the cloud-side feature blocks, the edge model extracts features based on the collaborative budget request and performs general training on the cloud-side feature blocks. This round of value evaluation returns the remaining collaborative budget to prepare for the next round of training. During each update, the cloud-side and edge models learn from each other to improve performance. Among them, P represents the pruning operation, Δ value Indicates the degree of pruning applied to the feature extraction block of the cloud-side model; The collaborative training process described above: During the cloud-to-edge process, the cloud-side feature extraction block is combined with the edge-side features and trained together with the original edge-side fully connected layer. During the cloud-to-edge collaborative training, the parameters of the cloud-side feature extraction block are fixed, and only the features of the edge-side model itself are updated. Throughout the entire process, the edge-side model absorbs the knowledge shared by the cloud side to achieve improved task granularity. Define a collaborative training convolutional network as: in, and They represent the cloud-side feature blocks and the edge-side feature blocks respectively, and ⊙ represents the feature collaborative training process.

5. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 1 is characterized by: The training focus of the edge-to-cloud process shifts from the edge to the cloud. Features are provided by the edge, and data is provided by the cloud. To ensure fair interaction between the cloud and edge, the cloud-side feature extraction module performs pruning in the cloud-to-edge process, while edge features are noise-added in the E2C process. The specific definitions are as follows: In the feature extraction module, the edge side will evaluate the cloud side feature extraction block through information gain The value coefficient θ j ; This value coefficient is used to determine the appropriate noise addition to the side model features to prevent feature leakage; in, is the side feature after noise processing, The mean is 0 and the standard deviation is Gaussian noise.

6. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 1 is characterized by: When the target is unknown, feature collaborative training based on the attention feature fusion mechanism is used to improve the granularity and accuracy of the edge model, while also improving the generalization and diversity of the cloud model. This is achieved through four parts: edge-cloud interconnection verification mechanism, cooperative budget update design, unknown target collaborative training scenarios, and reinforcement learning optimization. The edge-cloud interaction verification mechanism includes two key components: the cloud-edge interaction handshake protocol and the information gain evaluation mechanism.

7. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 6 is characterized by: The collaborative budget update design specifically uses a collaborative budget update method to perform collaborative training between the cloud side and the edge side while protecting the target privacy; Assuming there are n edge models, in the initial iteration of training, the cloud side will randomly initialize the collaboration budget of each edge model To meet the conditions Subsequently, based on the cloud-edge interaction handshake protocol, edge-cloud collaborative interaction is carried out, and each handshake process represents an interaction; Collaborative training between models is an iterative process that requires multiple rounds of interaction. Therefore, an edge-cloud coordination point is set up to achieve fairness and privacy protection in multiple rounds of collaborative interactive training. In the proposed edge-cloud collaborative training round, each edge feature interacts with its corresponding cloud-side feature extraction block to form a pair of edge-cloud feature pairs. The value evaluation coefficient θ of each pair of edge-cloud features is j Represents the degree of correlation between the two; The feedback mechanism introduced includes the reverse gradient balancing mechanism and the collaborative budget recovery mechanism; The reverse gradient balancing mechanism is based on the value evaluation coefficient θ of each edge model j Redistribute the collaboration budget; there are m edge models, and their corresponding value evaluation coefficients are θ j ,j=1,…,m; the cooperation budget will be based on these value assessment coefficients After training is completed, if there are m edge models waiting for collaboration budget reallocation, the cloud side will reallocate them based on the reallocation coefficient. Redistribute the collected collaboration budget in reverse order; The collaborative budget recovery mechanism is based on θ j The gradient value of the collaboration budget is recovered; if the θ of a certain edge model j The gradient change is close to zero, indicating that its performance improvement has reached its limit. At this time, the cloud side can partially reclaim the collaboration budget of the edge model and reallocate it to other edge models waiting for training; the edge model will retain part of the collaboration budget to continue collaborative training with other larger models.

8. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 6 is characterized by: The unknown target collaborative training scenario is specifically: Assume there are n edge features and one cloud side target; use θ1,θ2,…,θ n Represents the information gain coefficient of each side model feature, using a1, a2, ..., a n Represents its corresponding attention weight; the attention weight a of each side model feature j j The calculation formula is as follows: Given the feature maps of the cloud side and the edge side Introducing attention weights to achieve attention-based feature fusion between cloud-side models and edge models; in, Represents the fused features, C represents the number of channels, H represents the number of samples, and W represents the dimension of the features; Represents the attention weight generated by MS-CAM. Introducing a scalar value a j As the updated attention weight, where a j ∈[0,1] and 1-a j This constraint is also satisfied; the network can and Soft selection or weighted averaging between them.

9. The efficient federated collaborative learning framework for edge-cloud heterogeneous model co-evolution according to claim 6 is characterized by: The reinforcement learning optimization is specifically: Edge-cloud collaborative training is expressed as a Markov decision process, using the policy gradient π(a|λ c ,λ e ;ε,θ) to optimize the collaborative strategy; Among them, the policy gradient consists of two parts: the first part is: π c (i|λ c ; ε e ) According to the feature map λ in the cloud-side feature unit c Select the feature extraction block for co-training; the second part is: π e (j|λ e ; θ) is used to select appropriate end-side features for collaborative training; these two parts are integrated into a loss function for optimization: Part I, Strategy π c Defined by the cloud-side model prediction accuracy: Where i = 1,…,N, is the i-th score of the cloud-side collaborative training effect; the return rate during the holding period t is in is the rate of increase of prediction accuracy; given the initial score The cumulative return of a trajectory τ is: The optimization goal of this part is to maximize the logarithmic cumulative benefit of all trajectories τ: Part II, Strategy π e Defined by the edge-side model prediction accuracy: Where, j = 1,…,M, is the jth score of the edge-cloud collaborative training effect; similar to the cloud side, the return rate for holding period t is Given an initial score The cumulative gain of a trajectory is: Therefore, the optimization objective of the side model is: That is, the overall objective function is and The weighted sum of: Among them, ζ is a hyperparameter used to control the different learning rates of each part; these two parts are Optimize at the same time.

10. Application of an efficient federated collaborative learning framework for "edge-cloud" heterogeneous model co-evolution according to any one of claims 1 to 9 in medical image analysis, industrial Internet of Things, or autonomous driving.

Images