An individualized federated learning method for cloud-edge collaboration of the Internet of Things

Abstract

The application discloses a personalized federated learning method for cloud-edge collaboration of Internet of Things, which is characterized by resource limitation and high heterogeneity in the cloud-edge collaborative Internet of Things scene. Dynamic parameter decomposition and integration are introduced into personalized federated learning. In view of the potential risks of catastrophic forgetting and bias accumulation in the parameter transmission and aggregation process of federated learning, an adaptive regularization optimization method based on continuous learning is adopted. To solve the above problems, firstly, the global shared parameter and the client-specific parameter are decoupled by using the parameter decomposition strategy. Then, the local optimization process performed by each client is compared to sequential multitasking, and the online importance score and sensitivity in the optimization path are calculated. Finally, an elastic regularization term based on the learning trajectory is introduced to constrain the parameter update, and the learning rate of the next round is dynamically adjusted according to the importance weight of each parameter in the last round of global tasks. This method effectively balances the global sharing and local individualization requirements, and is particularly suitable for cloud-edge collaborative Internet of Things environments with limited resources, frequent communication and strong data heterogeneity. It not only improves the accuracy and robustness of the model on the edge side, but also significantly reduces the communication and computing overhead, and has good practical deployment prospects and expansion capabilities.

Description

Technical Field

[0001] This invention belongs to the field of federated learning technology, and specifically relates to a personalized federated learning method for cloud-edge collaboration in the Internet of Things. Background Technology

[0002] In IoT scenarios, the massive amounts of data generated by terminal devices (such as smart sensors and wearable devices) are characterized by discrete distribution and data privacy sensitivity. To fully utilize this data, secure methods are needed to achieve cross-domain collaborative training. Federated learning (FL), as a novel distributed training paradigm, fully utilizes the data collected by IoT terminals and enables resource-constrained edge devices to collaboratively train models without moving the original data. This effectively alleviates the data silo problem while protecting privacy, and has significant application potential in multiple fields.

[0003] In practical applications, different clients have different preferences for model performance. Personalized federated learning (PFL) optimization is an effective method to overcome the limitations of heterogeneous data and insufficient individual performance. Model-based PFL methods have two objectives: optimizing the global federated learning to meet potential downstream personalization needs, and improving personalization performance from the global federated learning model. However, existing personalization methods ignore the potential risks of catastrophic forgetting and bias in the global model, especially when frequent local updates are performed on edge devices. These problems will be further exacerbated, affecting the model's generalization ability and personalization effect. For example, model fine-tuning or multi-task learning either leads to excessively high communication overhead or fails to balance the trade-off between adaptability and stability.

[0004] The performance degradation is caused by client drift, which arises from inconsistent optimization objectives among edge clients leading to weight divergence, and the accumulation of gradient skew during multiple rounds of local update iterations. After training and updating model parameters for a new task, the old gradient update directions are forgotten, with only the new update task being considered while the previously learned old task is ignored. Traditional regularization methods in federated learning do not consider the evolution of parameter importance over time, potentially leading to catastrophic forgetting of gradients.

[0005] Therefore, it is urgent to address the challenge of balancing stability and personalization in non-independent and identically distributed (non-IID) data environments, effectively preventing catastrophic forgetting while ensuring high adaptability to personalized update needs. This paper proposes a strategy that leverages the coupling of global training and local fine-tuning to improve client performance, where data heterogeneity among clients stems from user-specific behaviors and geographical differences. Simultaneously, a federated optimization strategy is urgently needed that not only protects parameters critical to the global task but also provides greater flexibility for local updates on clients, enabling dynamic evaluation and retention of key global parameters, mitigating drift across heterogeneous data, and still achieving effective personalization for each client. Summary of the Invention

[0006] To address the above problems, this invention provides a personalized federated learning method based on continuous learning and dynamic parameter optimization, with the following specific steps:

[0007] S1. Dynamic Parameter Decomposition: The data collected by each edge device at the terminal is used as the local training set. In order to provide a suitable starting point in the early stage of training, a global model is initialized and parameter decomposition is performed. The parameter configuration of the loss sliding window is used to specify the timing of activating Dynamic Parameter-Aware Optimization (DPAO) and starting parameter decomposition.

[0008] S2. Each client uses local data collected by edge devices to train the local model based on local optimization objectives and updates the personalized parameters of the local model.

[0009] S3, Online Importance Score: Based on Shared Parameters of Client i The empirical Fisher Information Matrix (FIM) is used to continuously estimate the importance score F of each shared parameter using an efficient online resilient weight ensemble method. i This avoids costly recalculation for each task;

[0010] S4. Optimize path sensitivity: Based on F in step S2 i Calculate the importance score for each parameter in the training trajectory, estimate the optimized path integral by accumulating the contribution ratio of each update, and measure the dynamic path sensitivity score S(w) through a sliding time window in the most recent training step. i );

[0011] S5. Adaptive Weighted Aggregation: In each round of local update on client i, dynamic importance weights Ω are constructed based on the online FIM with moving average and path sensitivity accumulation, as described in step S3. i We propose an elastic regularization term based on the learning trajectory to constrain the global model parameter update and dynamically adjust the learning rate in the next round according to the importance weight of each parameter in the previous round of global task.

[0012] S6. Parameter Scoring Update: Introduce "peak-to-stable" detection on the server side, and update the client importance weight related parameters F stored on the server side during the stabilization period of loss changes. i ,S(w i ).

[0013] Furthermore, the dynamic parameter decomposition in step S1 mainly includes the following process:

[0014] S1.1 Assume that each client i uses an L-layer deep neural network model to perform a multi-class classification task, and the parameter set of this neural network is optimized by integrating all network layers k∈L:

[0015]

[0016] Among them, w i,k The specific parameters corresponding to the k-th layer of the local model held by client i;

[0017] S1.2 will assign each client parameter w i It is broken down into globally shared parameters (cross-client knowledge) and locally personalized parameters (client-specific knowledge), as follows:

[0018]

[0019] in, Globally shared parameters for each client It mainly abstracts the common knowledge among clients, while each client's personalized parameters are... Capture personalized knowledge;

[0020] S1.3 The parameters for configuring the loss sliding window are used to specify when to activate dynamic parameter-aware regularization and initiate parameter decomposition.

[0021] Furthermore, step S5, adaptive weighted aggregation, mainly includes the following processes:

[0022] S5.1 sets hyperparameters for global tasks and simultaneously enhances the flexibility of non-critical parameters.

[0023] S5.2 After combining the importance scores obtained from the Functional Independence Scale (FIM) with the optimization path, we use these scores as regularization constraints for the globally shared parameters:

[0024]

[0025] Where λ is the forgetting coefficient, F i (t-1) It is parameter w i The experience of FIM, and It is the path integral from the first communication round t0 to the last training iteration t-1, where t is the current communication round.

[0026] S5.3 Averaging the scores of the globally shared parameters after each task can reduce the sensitivity of the regularization hyperparameter to the number of tasks;

[0027] S5.4 Combining the EWC++ importance score calculated in step S2 and the optimized path sensitivity calculated in step S3, a weighted quadratic penalty is applied to each shared parameter to achieve a balance between stability and flexibility in each round.

[0028] S5.5 normalizes each parameter and automatically adjusts the constraint strength, balancing stability and individualization in the case of non-independent and identically distributed systems.

[0029] Furthermore, step S6, parameter score update, mainly includes the following process:

[0030] S6.1 Calculate the mean μ and standard deviation σ of the current window, set the smoothing window threshold T, and use a sliding window to detect the stabilization period of the loss;

[0031] S6.2 Once the model stabilizes, the continuous learning method updates its meta-knowledge to avoid forgetting previously learned tasks, and uses the period when the global loss first reaches its peak and then drops to a stable state as the time to update the importance weights.

[0032] S6.3 first performs peak detection; if μ cur ≥μ prev +σ cur If a peak is detected, initiate stability monitoring; if the detection reaches a stable period, if μ cur +σ cur If the value is less than T, it is considered that the new task has entered a learning plateau period after the peak, thus triggering a synchronous update of the global importance weight;

[0033] S6.4 broadcasts the updated global shared parameters to each edge server, awaiting the next round of personalized training and optimization aggregation.

[0034] This invention not only protects parameters critical to the global task but also provides greater flexibility for local updates on the client side. Furthermore, by designing an efficient importance weight update mechanism, it significantly reduces communication and computational overhead compared to existing methods.

[0035] 1. This invention integrates online importance estimation and path integral sensitivity scoring to design a task-independent dynamic regularization mechanism. The importance weights are updated through an event-triggered mechanism (loss stabilization period), thereby effectively reducing communication and computing overhead and making it more suitable for cloud-edge collaboration in resource-constrained IoT environments.

[0036] 2. This invention proposes a two-dimensional optimized adaptive federated learning method, which effectively separates the adaptation to specific client needs from the preservation of globally shared knowledge by utilizing parameter decomposition and a flexible weight integration strategy based on continuous learning. Attached Figure Description

[0037] Figure 1 This is the overall flowchart of the present invention.

[0038] Figure 2 This is a diagram illustrating the overall framework of the method of this invention. It shows the end-to-end dynamic parameter-aware optimization process in personalized federated learning, including initialization and transformation, parameter decomposition, online importance scoring, local training with dynamic parameter-aware optimization regularization, weighted aggregation, and synchronous updates triggered by peak stability. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0040] This invention provides a personalized federated learning method based on continuous learning and dynamic parameter optimization, the overall framework of which is as follows: Figure 1 As shown, the specific steps are as follows:

[0041] S1. Dynamic Parameter Decomposition: Data collected by each edge device at the terminal is used as the local training set. To provide a suitable starting point in the early stages of training, a global model is initialized and parameter decomposition is performed. The parameter configuration of the loss sliding window is used to specify when to activate dynamic parameter-aware optimization and initiate parameter decomposition. The specific steps are as follows:

[0042] S1.1 Assume that each client i uses an L-layer deep neural network model to perform a multi-class classification task, and the parameter set of this neural network is optimized by integrating all network layers k∈L:

[0043]

[0044] Among them, w i,k The model parameters corresponding to the k-th layer passed by client i;

[0045] S1.2 divides parameters into globally shared parameters (cross-client knowledge) and locally personalized parameters (client-specific knowledge), as follows:

[0046]

[0047] in, Globally shared parameters for each client It mainly abstracts the common knowledge among clients, while each client's personalized parameters... Capture unique knowledge;

[0048] S1.3 The parameters for configuring the loss sliding window are used to specify when to activate dynamic parameter-aware regularization and initiate parameter decomposition;

[0049] S2. Each client uses local data collected by edge devices to train the local model based on local optimization objectives and updates the personalized parameters of the local model. Perform mini-batch stochastic gradient descent using a local optimization objective, as follows:

[0050]

[0051] Where η is the local learning rate of client i, B is the local mini-batch stochastic gradient descent period, and t is the global communication round.

[0052] S3, Online Importance Score: Based on Shared Parameters of Client i Based on the empirical Fisher information matrix, an efficient online resilient weighting method is used to continuously estimate the importance score F of each shared parameter. i This avoids costly recalculation for each task. The specific steps are as follows:

[0053] S3.1 Definition

[0054] S3.2 Calculates importance weights using an efficient online elastic weight consolidation method for each client parameter w. i and output distribution p w (y|x), the diagonal approximation term of FIM is defined as:

[0055]

[0056] Where, a∈[0,1] and Online updates are implemented using the moving average of the current batch;

[0057] S4. Optimize path sensitivity: Based on F in step S3 i Calculate the importance score for each parameter in the training trajectory, estimate the optimized path integral by accumulating the contribution ratio of each update, and measure the dynamic path sensitivity score S(w) through a sliding time window in the most recent training step. i The specific steps are as follows:

[0058] S4.1 enhances FIM by systematically accumulating parameter importance scores throughout the training process, with the sensitivity of parameter importance measuring the efficiency of each update in improving the loss;

[0059] S4.2 Calculate the optimization score of this optimized path, which is the ratio of the change in loss to the change in the approximate KL divergence (measured using online FIM), and its form is defined as follows:

[0060]

[0061] Where, Δw i During the training of task i, the parameter w within period T i The change in ò is a very small value used to prevent division by zero;

[0062] S4.3 defines the importance score based on the old FIM parameters and the optimization path score within the update cycle as a unified per-weight regularization term. Then, standardization processing is performed.

[0063] S5. Adaptive Weighted Aggregation: In each round of local update on client i, dynamic importance weights Ω are constructed based on the online FIM with moving average and path sensitivity accumulation, as described in step S3. i We propose an elastic regularization term based on the learning trajectory to constrain the global model parameter updates, and dynamically adjust the learning rate in the next round according to the importance weight of each parameter in the previous round of the global task. The specific steps are as follows:

[0064] S5.1 sets hyperparameters for global tasks and simultaneously enhances the flexibility of non-critical parameters.

[0065] S5.2 After combining the importance scores obtained from the Functional Independence Scale (FIM) with the optimization path, we use these scores as regularization constraints for the globally shared parameters:

[0066]

[0067] Where λ is the forgetting coefficient, F i (t-1) It is parameter w i The experience of FIM, and It is the path integral from the first communication round t0 to the last training iteration t-1, where t is the current communication round.

[0068] S5.3 Averaging the scores of the globally shared parameters after each task can reduce the sensitivity of the regularization hyperparameter to the number of tasks;

[0069] S5.4 Combining the EWC++ importance score calculated in step S2 and the optimized path sensitivity calculated in step S3, a weighted quadratic penalty is applied to each shared parameter to achieve a balance between stability and flexibility in each round.

[0070] S5.5 normalizes each parameter and automatically adjusts the constraint strength, balancing stability and individualization in the case of non-independent and identically distributed systems.

[0071] S6. Parameter Scoring Update: Introduce "peak-to-stable" detection on the server side, and update the client importance weight related parameters F stored on the server side during the stabilization period of loss changes. i ,S(w i The specific steps are as follows:

[0072] S6.1 Calculate the mean μ and standard deviation σ of the current window, set the smoothing window threshold T, and use a sliding window to detect the stabilization period of the loss;

[0073] S6.2 Once the model stabilizes, the continuous learning method updates its meta-knowledge to avoid forgetting previously learned tasks, and uses the period when the global loss first reaches its peak and then drops to a stable state as the time to update the importance weights.

[0074] S6.3 first performs peak detection; if μ cur ≥μ prev +σ cur If a peak is detected, initiate stability monitoring; if the detection reaches a stable period, if μ cur +σ cur If the value is less than T, it is considered that the new task has entered a learning plateau after the peak, thus triggering a synchronous update of the global importance weight.

[0075] S6.4 broadcasts the updated global shared parameters to each edge server, awaiting the next round of personalized training and optimization aggregation.

Claims

1. A personalized federated learning method for Internet of Things cloud-edge collaboration, characterized in that, Comprising the following steps: S1, dynamic parameter decomposition: each edge device collects data collected at the terminal as a local training set, in order to provide a suitable starting point in the early training, initialize the global model and perform parameter decomposition, configure the parameters of the loss sliding window to specify the timing of dynamic parameter-aware optimization (DPAO) and start parameter decomposition; S2, personalized training: each client trains the local model using the local data collected by the edge device according to the local optimization objective, and updates the personalized parameters of the local model ; S3, online importance score: based on client's shared parameters Fisher Information Matrix (FIM) of the client's shared parameters, using an online resilient weighting integration method to continuously estimate the importance score of each shared parameter , thus avoiding expensive recomputations for each task;​ S4, Optimized path sensitivity: Estimate the optimized path integral by accumulating the proportion of contribution of each update, in the recent training steps, the dynamic path sensitivity integral measured by a sliding time window ; S5、Adaptive weighted aggregation: at the client In each round of local update, the dynamic importance weight is constructed by estimating the parameter importance score in step S3 and the path sensitivity score in step S4 An elastic regularization term based on learning trajectory is proposed to constrain the global model parameter update, and the learning rate of the next round is dynamically adjusted according to the importance weight of each parameter in the last round of global task; S6, Parameter score update: Introduce "peak-stable" detection on server side, update client importance weight related parameters stored on server side in stable period of loss change , .

2. The method of claim 1, wherein, for the dynamic parameter decomposition, the method is characterized by, Assuming each client All adopt one A deep neural network model with multiple layers is used to perform multi-class classification tasks. The parameter set of this neural network integrates all network layers. Optimize: ， wherein, corresponding to the client the first layer's model parameters.

3. The method of claim 1, wherein the dynamic parameter decomposition is characterized by, Balancing the contributions of global and local, not only meets the heterogeneous needs of different clients, but also guarantees the generalization ability of the global model, divides the parameters into global shared parameters of cross-client knowledge and local personalized parameters of client-specific knowledge, as follows: ， wherein, global shared parameters mainly abstracts common knowledge between clients, while individualized parameters of each client capture individualized knowledge; after completing local updates, global shared parameters are transmitted to the server for aggregation, transmission vectors and client-specific parameters are still stored locally, waiting for updates during subsequent local model training, and are available during the testing phase.

4. The method of claim 1, wherein the dynamic parameter decomposition is characterized by, Configure the parameters of the loss sliding window to specify the timing of dynamic parameter-aware optimization and start parameter decomposition, decompose the initial global model to provide a stable starting point for training, in addition, only transmit global parameters to reduce communication overhead, while minimizing training oscillation.

5. The personalized federated learning method for IoT cloud-edge collaboration according to claim 1, characterized in that, for online importance scoring, The Fisher information matrix quantifies the sensitivity of model output to parameter changes and is used to identify parameters important to old tasks, which can track the direction of parameters with large gradient amplitudes.

6. The method of claim 1, wherein the online importance score is determined based on a number of times a user has accessed a service provided by the service provider. Online elastic weight integration method is used to continuously estimate the importance of each shared parameter, while avoiding high repeated calculation cost.

7. The method of claim 1, wherein the path sensitivity score is optimized. Path sensitivity measures the efficiency of each update in reducing loss, which is enhanced by systematically accumulating parameter importance scores throughout the training process. Optimization path sensitivity measures the efficiency of each update in improving loss.

8. The personalized federated learning method for IoT cloud-edge collaboration according to claim 1, wherein the elastic regularization term is characterized in that, The period when the global loss first reaches a peak and then drops to a stable state is used as the timing of updating the importance weight, and "peak-stable" detection is introduced at the server end, which includes: 1) Calculate the mean of the current window and standard deviation , introduce a smoothing window threshold , use a sliding window to detect a stable period of loss; 2) Peak detection: If then a peak is detected and stationary monitoring is initiated. 3) Stationary phase detection: i.e. If the condition is met, it is considered that the new task has entered the learning plateau after the peak, triggering the synchronous update of the global importance weight.

9. The personalized federated learning method for IoT cloud-edge collaboration according to claim 1, wherein the elastic regularization term is characterized in that, Once the model stabilizes, the continuous learning method updates its meta-knowledge to avoid forgetting previously learned tasks, combines the importance score calculated in step S3 and the optimization path sensitivity calculated in step S4, and uses these scores as regularization constraints for global shared parameters. Weighted quadratic penalty is imposed on each shared parameter to balance stability and flexibility in each round: ， wherein, is a forgetting factor, is a parameter of the empirical Fisher information matrix, is a client cumulative score from the first communication to the last training iteration of training.

10. The personalized federated learning method for IoT cloud-edge collaboration according to claim 1, wherein the elastic regularization term is characterized in that, Set the hyperparameters for the global task and at the same time enhance the plasticity of unimportant parameters, average the scores after each task to reduce the sensitivity of the regularization hyperparameters to the number of tasks, normalize each parameter and automatically adjust the constraint strength, and balance stability and individualization in the case of dependent and identically distributed.

Images