Federal learning tumble detection system based on adaptive weight optimization and differential privacy protection

Abstract

The invention discloses a federal learning tumble detection system based on adaptive weight optimization and differential privacy protection. The federal learning tumble detection system comprises a server module and a plurality of client modules in communication connection with the server module, the server module comprises a global model management unit which is used for initializing and storing a double-flow space-time diagram convolutional network as a global model; the communication interface unit is used for distributing the global model of the global model management unit to each client module and receiving privacy-protected local model update, corresponding local verification accuracy and accounting differential privacy consumption uploaded by each client module; and the self-adaptive weight calculation unit is used for verifying the accuracy rate according to each local received by the communication interface unit. According to the invention, on the premise of strictly protecting local data privacy of a user, a federated learning system which effectively improves fall detection precision and generalization ability is provided, so that the problem that privacy security and model performance are difficult to consider at the same time is solved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of federated learning, privacy computing, and human motion recognition, specifically to a federated learning fall detection system with adaptive weight optimization and differential privacy protection. Background Technology

[0002] With the accelerating aging of society and the increasing demands for public safety, high-precision fall detection technology has significant application value in areas such as smart elderly care and community security. Current mainstream fall detection solutions generally adopt a "centralized data collection + unified model training" approach, which has the following inherent drawbacks: First, there is a high risk of data privacy breaches: User skeletal motion data and posture trajectories are sensitive biometric information, which faces the risk of leakage during centralized storage and cross-network transmission, making it difficult to meet the compliance requirements of laws such as the Personal Information Protection Law and the Data Security Law. Second, the model's generalization performance is limited: the distribution of motion data varies significantly across different application scenarios such as home monitoring, community security, and medical institutions. Models trained in a centralized manner cannot effectively adapt to the data feature distribution of multiple scenarios, leading to a decrease in detection accuracy in actual deployment. To address these privacy protection challenges, federated learning technology has been introduced into the fall detection field in recent years. Through the model of "data not leaving the local environment and model co-evolving," it theoretically solves the privacy risks brought about by data centralization. However, existing federated learning fall detection schemes still suffer from the following technical bottlenecks: 1. Lack of adaptability in model aggregation strategies: Existing schemes mostly adopt fixed weight allocation or simple equal weight aggregation strategies based on data volume, failing to fully consider the differences in the contribution of local model performance of each client to the global model, thus restricting further improvement in model accuracy; 2. Insufficient dynamism in privacy protection mechanisms: Most schemes adopt fixed-intensity noise injection mechanisms, lacking the ability to dynamically adjust the privacy protection intensity according to the training stage and model state, making it difficult to achieve an optimal balance between privacy protection and model usability; 3. Limited action feature representation capabilities: Existing methods are mostly based on single-modal action feature extraction architectures, failing to fully explore the dual features of fall actions in spatial structure and temporal motion, thus limiting detection performance in complex scenarios.

[0003] Therefore, there is an urgent need to develop a federated learning fall detection system that integrates intelligent weight allocation, dynamic privacy protection, and multimodal feature extraction, so as to fundamentally solve the technical problem of balancing privacy and detection accuracy in existing technologies. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the present invention aims to provide a federated learning fall detection system with adaptive weight optimization and differential privacy protection. This federated learning system can effectively improve the accuracy and generalization ability of fall detection through adaptive weight aggregation and multimodal feature extraction while strictly protecting the privacy of users' local data, thus solving the problem of balancing privacy and model performance.

[0005] To achieve the above objectives, the following solution is adopted: A federated learning fall detection system with adaptive weight optimization and differential privacy protection includes a server module and multiple client modules that are communicatively connected to the server module. The server module includes: The global model management unit is used to initialize and store a two-stream spatiotemporal graph convolutional network as a global model. The communication interface unit is used to distribute the global model of the global model management unit to each client module, and to receive the privacy-protected local model update, the corresponding local verification accuracy, and the calculated differential privacy consumption uploaded by each client module. An adaptive weight calculation unit is used to calculate the adaptive aggregate weight corresponding to each client module based on the local verification accuracy received by the communication interface unit. The model aggregation unit is used to use the adaptive aggregation weights obtained by the adaptive weight calculation unit to perform weighted aggregation on each privacy-protected local model update received by the communication interface unit, generate updated global model parameters, and send the updated global model parameters to the global model management unit to update the global model. The privacy monitoring unit is used to receive and accumulate the differential privacy consumption reported by each client module from the communication interface unit, and trigger a termination signal when the total consumption reaches a preset privacy budget threshold. The client module includes: The data preprocessing unit is used to perform hierarchical sampling on the locally stored action dataset, dividing it into a local training set and a local validation set. The model training unit is used to train the global model downloaded through the communication unit and the local training set, generate an initial local model update, and calculate the local validation accuracy using the local validation set. The differential privacy protection unit is used to perform adaptive gradient clipping and dynamic Gaussian noise addition sequentially on the initial local model update generated by the model training unit to generate the privacy-protected local model update, and calculate the differential privacy consumption of this training. The privacy-protected local model update and the local verification accuracy are uploaded through the communication unit. The communication unit is used to download the global model from the server module and upload the differential privacy consumption generated by the differential privacy protection unit, the privacy-protected local model update, and the local verification accuracy calculated by the model training unit to the server module.

[0006] Furthermore, the dual-stream spatiotemporal graph convolutional network maintained by the global model management unit includes a spatial flow branch and a temporal flow branch; the spatial flow branch is used to extract the spatial structure features of the action data, and the temporal flow branch is used to extract the temporal motion features of the action data; the network fuses the spatial structure features and the temporal motion features by fusing the weight matrix and the bias term.

[0007] Furthermore, the adaptive weight calculation unit employs a softmax temperature coefficient optimization strategy, calculating the first weight according to the following formula. Adaptive aggregation weights for each client : in, For the first The accuracy of local verification received by each client module, where N is the number of client modules. This is the temperature coefficient, with a value ranging from 0.1 to 0.5.

[0008] Furthermore, the adaptive gradient pruning operation performed by the differential privacy protection unit is carried out according to the following formula: Where g is the model gradient and C is the clipping threshold, ranging from 0.2 to 0.3. Let g be the norm of the gradient g.

[0009] Furthermore, the dynamic Gaussian noise addition operation performed by the differential privacy protection unit is carried out according to the following formula: in, The model parameters after adding noise. These are the original model parameters, where N is a Gaussian distribution. Based on the standard deviation of the basic noise, Let I be the clipping threshold, and let I be the identity matrix.

[0010] Furthermore, the model aggregation unit performs weighted aggregation of local model updates according to the following formula to generate updated global model parameters: in, These are the global model parameters updated in round t+1. For from the first The client module receives the privacy-preserving local model update parameters for the t-th round. For the first The adaptive aggregation weight of each client module, where N is the total number of client modules.

[0011] Furthermore, the data preprocessing unit performs hierarchical sampling and partitioning of the local dataset according to the following formula: in, For the first Local action datasets for each client module. The ratio of the training set to the validation set is: .

[0012] Furthermore, the model training unit employs two-layer Dropout regularization during training, wherein the first layer Dropout probability is 0.3 and the second layer Dropout probability is 0.2.

[0013] Furthermore, the target privacy budget set by the privacy monitoring unit is ε ≤ 10, δ = 1e-4.

[0014] Furthermore, the privacy monitoring unit will also use reaching a preset number of training rounds as a condition to trigger the termination signal; after the training is terminated, the final global model maintained by the global model management unit is used to identify and detect multiple types of human actions, including falling actions, on the client module.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively improves the accuracy and cross-scenario generalization ability of the federated learning global model. The invention employs a server-side softmax temperature coefficient adaptive weight optimization strategy, dynamically calculating aggregate weights based on the client-side local validation accuracy, and then performing weighted aggregation accordingly. This mechanism overcomes the model bias problem caused by non-independent and identically distributed (Non-IID) data in traditional solutions. By assigning higher aggregation weights to clients with better performance, it guides the global model towards a more accurate and generalizable direction, thereby significantly improving the recognition accuracy and generalization ability of the fall detection model in multiple scenarios.

[0016] 2. This invention constructs a rigorous, dynamic, and provable privacy protection system to ensure the security of sensitive data. By implementing a three-tiered privacy protection mechanism on the client side—adaptive gradient pruning, dynamic Gaussian noise generation, and Rényi differential privacy calculation—this invention achieves refined and dynamic protection of the local training process. This mechanism can adjust the protection strength according to the training status and strictly monitor the accumulated consumption on the server side through a privacy monitoring unit, ensuring that the entire training process meets (…). , This is conducted under strict privacy constraints. The entire system ensures that sensitive biometric information such as users' skeletal motion data does not leave the user's local machine, fundamentally avoiding the risk of privacy leaks and complying with relevant data security regulations.

[0017] 3. This invention significantly enhances the model's ability to characterize and recognize complex fall movements. The invention employs a dual-stream spatiotemporal graph convolutional network as the foundational architecture of the global model. This network extracts spatial structural features (skeletal joints) of the movement through a spatial flow branch and temporal motion features (movement sequences) through a temporal flow branch, and then deeply fuses the two. This design enables the model to gain a deep understanding of fall behavior from both static posture and dynamic process dimensions, solving the problem of insufficient characterization capabilities of traditional single-feature extraction architectures, thereby significantly improving the accuracy of recognizing various complex and unconventional fall movements.

[0018] In summary, this invention successfully integrates core technologies such as dual-stream spatiotemporal graph convolutional networks, adaptive weight aggregation based on softmax temperature coefficients, and a three-level differential privacy protection and global budget monitoring system ("pruning-noise addition-calculation") to achieve federated learning. In this model, the raw action data from each client module is processed and trained locally without being uploaded to the server module. The server module only aggregates the differentially privacy-protected model parameter updates uploaded by each client module. This allows for joint training of a high-precision, highly generalized fall detection model while meeting strict privacy requirements, fundamentally solving the technical challenge of balancing privacy and detection accuracy. Attached Figure Description

[0019] Figure 1 This is a block diagram of the federated learning fall detection system with adaptive weight optimization and differential privacy protection in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the three-level privacy protection process of the client module in this embodiment of the invention. Figure 3 This is a flowchart of the adaptive weight aggregation and model iteration process of the server module in this embodiment of the invention. Detailed Implementation

[0020] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0021] This invention provides a federated learning fall detection system with adaptive weight optimization and differential privacy protection, which is particularly suitable for scenarios with strict requirements for data privacy protection, such as elderly care, community safety protection, and medical rehabilitation monitoring.

[0022] like Figure 1 As shown, the adaptive weight optimization and differential privacy-preserving federated learning fall detection system of this invention includes a server module and multiple client modules that are communicatively connected to the server module.

[0023] The server module includes: The global model management unit is used to initialize and store a two-stream spatiotemporal graph convolutional network as a global model.

[0024] The communication interface unit is used to distribute the global model of the global model management unit to each client module, and to receive the privacy-protected local model updates, the corresponding local verification accuracy, and the calculated differential privacy consumption uploaded by each client module.

[0025] An adaptive weight calculation unit is used to calculate the adaptive aggregate weight corresponding to each client module based on the local verification accuracy received by the communication interface unit.

[0026] In other embodiments, the adaptive weight calculation unit may be used to calculate the adaptive aggregate weight corresponding to each client module based on other performance indicators received by the communication interface unit based on the client module uploads.

[0027] The model aggregation unit is used to perform weighted aggregation on each privacy-protected local model update received by the communication interface unit using the adaptive aggregation weight obtained by the adaptive weight calculation unit, generate updated global model parameters, and send the updated global model parameters to the global model management unit to update the global model.

[0028] The privacy monitoring unit is used to receive differential privacy consumption reported by each client module from the communication interface unit and accumulate it, and trigger a termination signal when the total consumption reaches a preset privacy budget threshold.

[0029] The client module includes: The data preprocessing unit is used to perform hierarchical sampling on the locally stored action dataset, dividing it into a local training set and a local validation set.

[0030] The model training unit is used to train the global model downloaded through the communication unit and the local training set, generate an initial local model update, and calculate the local validation accuracy using the local validation set.

[0031] The differential privacy protection unit is used to perform adaptive gradient clipping and dynamic Gaussian noise addition sequentially on the initial local model update generated by the model training unit to generate the privacy-protected local model update, and calculate the differential privacy consumption of this training. The privacy-protected local model update and the local verification accuracy are uploaded through the communication unit.

[0032] The communication unit is used to download the global model from the server module and upload the differential privacy consumption generated by the differential privacy protection unit, the privacy-protected local model update, and the local verification accuracy calculated by the model training unit to the server module.

[0033] Preferably, the adaptive weight calculation unit adopts a softmax temperature coefficient optimization strategy, which calculates the aggregate weight of each client module by softening the performance indicators of the client modules. The calculation formula is as follows: in, Let be the performance metric (verification accuracy) of the i-th client module. For the number of client modules, This is the temperature coefficient, with a value ranging from 0.1 to 0.5.

[0034] Preferably, the differential privacy protection unit implements a three-level differential privacy mechanism, including: adaptive gradient pruning, dynamic Gaussian noise addition, and Rényi differential privacy calculation; The adaptive gradient clipping formula is as follows: in, For the model gradient, The clipping threshold (values ​​from 0.2 to 0.3). The gradient norm is used; the formula for adding dynamic Gaussian noise is: in, The model parameters after adding noise. These are the original model parameters. Based on the standard deviation of the basic noise, This is the clipping threshold; The Rényi differential privacy accounting formula is as follows: in, It is the Rényi order. This is a privacy parameter, with a value of 1e-4.

[0035] Preferably, the global model is a two-stream spatiotemporal graph convolutional network, including a spatial flow branch and a temporal flow branch, which are used to extract the spatial structure features and temporal motion features of the action, respectively. Its feature fusion formula is as follows: in, Characteristics of spatial flow As a time-flow feature, To fuse the weight matrix, This is a bias term.

[0036] Preferably, the model aggregation unit performs weighted aggregation of model updates from each client module according to aggregation weights, and the aggregation formula is as follows: in, For the global model parameters in round t+1, Let be the local model parameters of the i-th client module in round t. Let be the aggregate weight of the i-th client module.

[0037] Preferably, the data preprocessing unit is used to perform stratified sampling on the local action dataset, and the sampling formula is: in, For the local dataset of the i-th client, The optimal ratio for dividing the training set / validation set is 0.7:0.3.

[0038] Preferably, the model training unit employs two-layer Dropout regularization, with the first layer having a Dropout probability of 0.3 and the second layer having a Dropout probability of 0.2.

[0039] Preferably, the target privacy budget set by the privacy monitoring unit is ε≤10, δ=1e-4.

[0040] The detection method of the adaptive weight optimization and differential privacy-preserving federated learning fall detection system according to embodiments of the present invention includes the following steps: S1. The server initializes a two-stream spatiotemporal graph convolutional network as the global model; S2. Each client downloads the global model from the server and performs hierarchical sampling on the locally stored action dataset, dividing it into a local training set and a local validation set; S3. Each client uses the downloaded global model and the local training set to train, generates an initial local model update, and uses the local validation set to calculate the local validation accuracy. S4. Each client sequentially performs adaptive gradient clipping and dynamic Gaussian noise addition on the initial local model update to generate a privacy-preserving local model update, and calculates the differential privacy consumption of this training. S5. Each client uploads the privacy-protected local model update, the local verification accuracy, and the differential privacy consumption to the server; S6. The server calculates the adaptive aggregation weight for each client based on the local verification accuracy received from each client; S7. The server uses the adaptive aggregation weight to perform weighted aggregation on the privacy-preserving local model updates uploaded by each client, generates updated global model parameters, and updates the global model accordingly. S8. The server accumulates the differential privacy consumption reported by each client and determines whether the total consumption has reached the preset privacy budget threshold. If it has not reached the threshold and other termination conditions are not met, the server returns to step S2 to start the next round of training.

[0041] This invention discloses a federated learning fall detection system with adaptive weight optimization and differential privacy protection. The system employs a federated learning framework and is specifically applied to the training and updating of action recognition models. The system includes a client module and a server module. The client module uses a three-level mechanism—adaptive gradient pruning, dynamic Gaussian noise generation, and Rényi differential privacy accounting—to protect the privacy of local sensitive action data and completes local training based on a spatiotemporal graph convolutional network. The server module calculates adaptive aggregation weights based on the verification accuracy uploaded by each client using a softmax temperature coefficient optimization strategy. This weighted aggregation of model updates from each client generates a global model, and continuously monitors privacy budget consumption. This system adopts a federated learning paradigm. The original action data from each client module is processed and used for model training locally without being uploaded to the server module. The server module only aggregates the differentially privacy-protected model parameter updates uploaded by each client module, thereby achieving joint training. During training, strict differential privacy constraints are met: ε≤10, δ=1e-4. This makes the system suitable for scenarios with high data privacy requirements, such as elderly care and community security.

[0042] The detailed deployment configuration and execution flow of the adaptive weight optimization and differential privacy-preserving federated learning fall detection system according to an embodiment of the present invention are as follows: System deployment and configuration: Both the server and client run on Windows 11, using Python 3.10.17 and PyTorch 2.5.1. The global model is a dual-stream spatiotemporal graph convolutional network (Dual-stream ST-GCN) with a spatial graph convolution strategy set to 'spatial', performing an 8-class classification task.

[0043] System initialization: like Figure 1 As shown, the adaptive weight optimization and differential privacy-preserving federated learning fall detection system of this invention adopts a client / server federated learning architecture, including one server module and multiple (four in this example) client modules. The server module initializes the global model. Each client module connects to the server and downloads the initial model via the communication interface unit and the communication unit. The server privacy monitoring unit is set to a target privacy budget ε≤10.0, δ=1e-4. The training epochs are set to 8, and the number of clients is [not specified]. There are 4.

[0044] Local training process on the client: 1. Data Preprocessing: The client-side data preprocessing unit loads the local action dataset (e.g., skeletal node coordinate features stored in pickle format), containing feature data and label data. The feature data is a four-dimensional tensor with dimensions [number of samples, time steps, number of joints, coordinate dimension]. It is adjusted to [number of samples, coordinate dimension, time steps, number of joints] using the permute(0, 3, 1, 2) operation to fit the model input. There are eight action categories: standing, walking, sitting, lying down, standing up, sitting down, and falling. Stratified sampling is used to divide the training and validation sets in a 7:3 ratio (ratio=0.7:0.3), with a random seed of random_state=42. Stratification is based on the argmax distribution of the labels to ensure a balanced distribution across categories.

[0045] 2. Model Training: The client-side model training unit employs a two-layer Dropout regularization, with a first-layer Dropout probability of 0.3 and a second-layer Dropout probability of 0.2. The AdamW optimizer is used with a learning rate of 0.00035 and weight decay of 5e-6. The loss function is cross-entropy loss. The ReduceLROnPlateau learning rate scheduler is configured; if the validation loss shows no improvement for two consecutive rounds, the learning rate decays to 0.5 of its original value. Motion vectors are calculated for each batch: mot = pts[:, :2, 1:, :] - pts[:, :2, :-1, :], inputting the spatial and temporal branches respectively.

[0046] 3. Privacy protection measures: such as Figure 2 As shown, the differential privacy protection unit performs three levels of processing sequentially during local training: (a) Adaptive gradient clipping: After the model gradient is generated, the L2 norm of the gradient is first limited by gradient clipping (example parameter C=0.25) to constrain the maximum impact of a single sample on the model.

[0047] (b) Dynamic Gaussian noise addition: Preferably starting from the second round of training, Gaussian noise is added to the clipped gradients (or parameters), with a noise standard deviation of [missing value]. Dynamically adjusted based on training rounds.

[0048] (c) Rényi Privacy Budget Accounting: Employing the Rényi Differential Privacy Accounting (RDP) mechanism ( The sampling value ranges from 1.1 to 64.0, and privacy consumption is calculated in real time and rigorously. Ensure that total consumption does not exceed the threshold.

[0049] Server aggregation process: like Figure 3 As shown, the aggregation process of the server module includes: 1. Weight Calculation: The server's adaptive weight calculation unit receives the verification accuracy data uploaded by each client. The softmax temperature coefficient strategy is adopted (example parameters). =0.3) Calculate the aggregate weight Preferably, a smooth update mechanism (smoothing coefficient) is adopted. Ensure weight stability.

[0050] 2. Model Aggregation: Model aggregation units are aggregated according to the calculated weights. Update the client model Weighted aggregation is performed, and the aggregation formula is: When aggregating, it is preferable to exclude integer type parameters (such as num_batches_tracked) to ensure the integrity of the model structure.

[0051] 3. Privacy Monitoring: The privacy monitoring unit accumulates the privacy budget for each client, and the total privacy consumption... Save the model checkpoint when ε≤10.0 and determine whether the termination condition is met (such as reaching the maximum training epoch). Real-time monitoring ensures that privacy constraints are met.

[0052] Training termination and model deployment: Training terminates when the preset number of training epochs (8 epochs) is reached or the privacy budget threshold is reached. The server saves the final global model. The client can download the final model for local fall detection tasks, supporting the recognition of 8 types of actions.

[0053] The federated learning fall detection system with adaptive weight optimization and differential privacy protection in this invention improves the model's accuracy on non-independent and identically distributed data by adaptive weight optimization, enhances feature extraction capabilities by dual-stream spatiotemporal graph convolution, and ensures privacy security through a three-level differential privacy mechanism, thus achieving a good balance between privacy protection and detection accuracy.

[0054] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A federated learning fall detection system with adaptive weight optimization and differential privacy protection, characterized in that, The server module and a plurality of client modules connected in communication with the server module; The server module comprises: a global model management unit for initializing and storing a dual-flow spatio-temporal graph convolution network as a global model; a communication interface unit for distributing the global model of the global model management unit to each client module, and receiving the privacy-protected local model updates uploaded by each client module, the corresponding local verification accuracy, and the calculated differential privacy consumption; an adaptive weight calculation unit for calculating the adaptive aggregation weight corresponding to each client module according to the local verification accuracy received by the communication interface unit; a model aggregation unit for weighting and aggregating each privacy-protected local model update received by the communication interface unit using the adaptive aggregation weight obtained by the adaptive weight calculation unit, generating updated global model parameters, and sending the updated global model parameters to the global model management unit to update the global model; a privacy monitoring unit for receiving and accumulating the differential privacy consumption reported by each client module from the communication interface unit, and triggering a termination signal when the total consumption reaches a preset privacy budget threshold; The client module comprises: a data preprocessing unit for stratified sampling of a locally stored action data set, dividing it into a local training set and a local validation set; a model training unit for training the global model downloaded through the communication unit with the local training set to generate an initial local model update, and calculating the local verification accuracy using the local validation set; a differential privacy protection unit for sequentially performing adaptive gradient clipping and dynamic Gaussian noise addition on the initial local model update generated by the model training unit to generate the privacy-protected local model update, and calculating the differential privacy consumption of this training, the privacy-protected local model update and the local verification accuracy being uploaded through the communication unit; a communication unit for downloading the global model from the server module, and uploading the calculated differential privacy consumption, privacy-protected local model update and local verification accuracy calculated by the model training unit generated by the differential privacy protection unit to the server module.

2. The federated learning fall detection system of claim 1, wherein, The dual-flow spatio-temporal graph convolution network maintained by the global model management unit comprises a spatial flow branch and a temporal flow branch; the spatial flow branch is used to extract the spatial structure features of the action data, and the temporal flow branch is used to extract the time motion features of the action data; The network fuses the spatial structure features and the time motion features through a fusion weight matrix and a bias term.

3. The federated learning fall detection system of claim 1, wherein, The adaptive weight calculation unit adopts a softmax temperature coefficient optimization strategy, and calculates the adaptive aggregation weight of the i-th client according to the following formula : w i = e x p ( s i / T ) / Σ j e x p ( s j / T ) : wherein, is the local verification accuracy received by the i-th client module, N is the number of client modules, is the local verification accuracy received by the i-th client module, N is the number of client modules, is the temperature coefficient, and the value range is 0.1 to 0.

5.

4. The federated learning fall detection system of adaptive weight optimization and differential privacy protection according to claim 1, wherein, The adaptive gradient clipping operation performed by the differential privacy protection unit is performed according to the following formula: where g is the model gradient, C is the clipping threshold, and the value range is 0.2 to 0.3, is the norm of the gradient g.

5. The federated learning fall detection system of adaptive weight optimization and differential privacy protection according to claim 4, wherein, The dynamic Gaussian noise addition operation performed by the differential privacy protection unit is performed according to the following formula: wherein, is the model parameter after adding noise, is the original model parameter, N is a Gaussian distribution, is the standard deviation of the basic noise, is the clipping threshold, I is an identity matrix.

6. The federated learning fall detection system of adaptive weight optimization with differential privacy protection according to claim 1, wherein, The model aggregation unit weights and aggregates the local model updates according to the following formula to generate updated global model parameters: wherein, is the updated global model parameter for the t+1th round, is the tth round privacy-protected local model update parameter received from the i th client module, is the adaptive aggregation weight for the i th client module, and N is the total number of client modules.

7. The federated learning fall detection system of adaptive weight optimization with differential privacy protection according to claim 1, wherein, The data preprocessing unit stratifies and samples the local data set according to the following formula: wherein, is the local action dataset for the client module, is the training set and validation set split ratio .

8. The federated learning fall detection system of adaptive weight optimization with differential privacy protection according to claim 1, wherein, The model training unit adopts double-layer Dropout regularization in the training process, wherein the first layer Dropout probability is 0.3, and the second layer Dropout probability is 0.

2.

9. The federated learning fall detection system of claim 1, wherein, The privacy monitoring unit sets the target privacy budget as epsilon <= 10, delta = 1e-4.

10. The federated learning fall detection system of adaptive weight optimization with differential privacy protection according to claim 1, wherein, The privacy monitoring unit also sets reaching the preset training round as a condition for triggering the termination signal; after the training is terminated, the final global model maintained by the global model management unit is used for recognizing and detecting multiple human body actions including falling actions on the client module.

Images