Federal hierarchical learning method based on differential privacy

Abstract

The invention relates to the technical field of federal learning, in particular to a federal hierarchical learning method based on differential privacy, and the method specifically comprises the following steps: 1, designing a system architecture, and constructing a multi-layer federal hierarchical learning architecture which comprises a bottommost layer, a middle layer and a topmost layer; step 2, local model training: at the bottommost layer, each data owner i has a local data set Di, the number of samples is Di, and the data owner i uses a stochastic gradient descent optimization algorithm to train a local model # imgabs0 # based on the local data set Di, and theta i is a model parameter; according to the method, differential privacy noise is added to the nodes of the bottommost layer and the middle layer of privacy protection, dynamic privacy budget allocation is combined, a multi-layer protection barrier is constructed, privacy leakage in the model parameter transmission process is prevented, original data safety is ensured, and privacy data leakage is avoided.

Description

Technical Field

[0001] The present invention relates to the field of federated learning technology, and in particular to a federated hierarchical learning method based on differential privacy. Background Art

[0002] In today's big data era, the value of data is becoming increasingly prominent. Machine learning, as a key technology for extracting information from data and enabling intelligent decision-making, has gained widespread application. However, with increasing attention to data privacy and security, traditional centralized machine learning faces numerous challenges, such as data silos and data compliance issues. Federated learning has emerged as a key technology. It allows multiple participants to collaboratively train machine learning models without sharing the original data, effectively resolving the data silo issue while also protecting data privacy to a certain extent.

[0003] Federated hierarchical learning is an extension of federated learning. It organizes multiple participants into a hierarchical structure, performing model training and parameter aggregation at different levels. This improves training efficiency and model performance, making it particularly suitable for large-scale, distributed scenarios. However, in federated hierarchical learning, while raw data is not directly shared, the risk of privacy leakage still exists during model training and parameter transmission. For example, malicious actors could infer sensitive data of other participants by analyzing model parameter updates. Summary of the Invention

[0004] In response to the technical problem in the existing technology that there is still a risk of privacy leakage during model training and parameter transmission, the present invention provides a federated hierarchical learning method based on differential privacy.

[0005] The technical solution adopted by the present invention is: a federated hierarchical learning method based on differential privacy, which specifically includes the following steps:

[0006] Step 1: System architecture design: building a multi-layer federated hierarchical learning architecture, including the bottom layer, middle layer, and top layer;

[0007] Step 2: Local model training. At the bottom layer, each data owner i has a local dataset D i , the number of samples is |D i |, data owner i based on local dataset D i , use the stochastic gradient descent optimization algorithm to train the local model where θ i are model parameters;

[0008] Step 3: Add differential privacy noise. Before uploading the model parameters to the bottom-level node, add differential privacy noise to it.

[0009] Step 4: Parameter aggregation from the bottom layer to the middle layer. The middle layer node j receives the model parameter update amount after adding noise from multiple nodes in the bottom layer and performs aggregation operation.

[0010] Step 5: Add differential privacy noise to the middle layer. Before the middle layer node uploads the aggregated parameter update to the higher layer, differential privacy noise is added again.

[0011] Step 6: Aggregate and transmit parameters between the middle layers;

[0012] Step 7: Update the central server model;

[0013] Step 8: Perform global model distribution;

[0014] Step 9: Allocate privacy budget;

[0015] Step 10: Model performance evaluation and optimization. During the training process, the performance of the global model is evaluated regularly.

[0016] In the above technical solution, at the architectural level, a multi-layer federated hierarchical learning architecture organizes distributed participants in an orderly manner. Aggregation nodes in the middle layer reduce data transmission volume and alleviate pressure on central servers, laying the foundation for privacy protection and parameter aggregation, addressing data silos and efficiency issues. Regarding privacy protection, differential privacy noise is added to the bottom and middle layer nodes. Combined with dynamic privacy budget allocation, this creates a multi-layered protection barrier to prevent privacy leakage during model parameter transmission, ensuring the security of original data and avoiding private data leaks. Furthermore, noise intensity and budget allocation are flexibly adjusted based on the scenario and training stage to balance privacy and performance. Regarding training optimization, local model training leverages data value and mitigates privacy risks; weighted average aggregation considers data volume differences and improves model stability; central server updates and global model distribution promote collaborative evolution; and regular performance evaluation and hyperparameter adjustments ensure continuous model optimization, improving accuracy and generalization to meet the needs of large-scale distributed scenarios.

[0017] In one embodiment, in step 1, the specific method for designing the system architecture and building a multi-layer federated hierarchical learning architecture is as follows:

[0018] First, suppose there are L layers in the system, with the bottom layer being l = 1, containing n1 data owners;

[0019] Among them, the middle layer is l=2,…,L-1 layer, and each layer contains a different number of aggregation nodes;

[0020] Finally, the top layer is l=L layer, which has a central server and uses N iRepresents the number of nodes in layer l. For node i in layer l, it receives model parameters from nodes in layer l-1 and uploads the aggregated model parameters to nodes in layer l+1.

[0021] In one embodiment, in step 2, in the tth round of training, the local model parameter update formula is:

[0022]

[0023] Where η is the learning rate, Indicates that based on the local dataset D i Calculated loss function About the model parameter θ i The gradient of , (x, y) is the sample and label in the dataset.

[0024] In one embodiment, in step 3, according to the Laplace mechanism, the model parameter update amount of the bottom node i is Add noise z i , get the update amount after adding noise

[0025] in,

[0026]

[0027] Among them, z i It follows the Laplace distribution The noise vector is Δf, which is the sensitivity (here refers to the maximum change in the amount of model parameter update). ∈ is the privacy budget, which controls the strength of privacy protection. The smaller ∈ is, the higher the privacy protection is, but it may have a greater impact on model performance.

[0028] In one embodiment, in step 4, a weighted average aggregation method is used, and the updated value of the model parameter after aggregation Δθ j for:

[0029]

[0030] Among them, S j Indicates the bottom-level node set that uploads parameters to the middle-level node j.

[0031] In one embodiment, in step 5, the aggregation parameter update amount Δθ for the intermediate layer node j is j , add noise z j , get the update amount after adding noise

[0032] in,

[0033] The z herej Also follows the Laplace distribution

[0034] Where Δf′ is the sensitivity of the update amount of the intermediate layer aggregation parameters, and ∈′ is the privacy budget of the intermediate layer.

[0035] In one embodiment, in step 6, for the middle layers where l=3, ..., L-1, the operations of steps 4 and 5 are repeated.

[0036] In one embodiment, the specific method for updating the central server model is as follows:

[0037] First, the central server (l = L) receives the aggregate parameter update after adding noise from the top intermediate nodes Update the global model parameters θ:

[0038]

[0039] Where T represents the set of top-level intermediate nodes that upload parameters to the central server.

[0040] In one embodiment, the specific method for performing global model distribution is as follows:

[0041] Update the global model parameters θ on the central server t+1 After that, the updated global model is distributed to nodes at each layer;

[0042] After receiving the global model, each layer of nodes uses it as the initial model for the next round of local model training and continues the local model training process in step 2, forming a closed-loop federated hierarchical learning process.

[0043] In one embodiment, the specific method for allocating the privacy budget is as follows:

[0044] Let the total privacy budget ∈ total Assign to different layers according to their importance and privacy risk level;

[0045] Assume that the privacy budget allocated to the lth layer is ∈ l ,satisfy

[0046] For the lowest-level nodes, since they are in direct contact with the original data, the privacy risk is higher, and a relatively larger privacy budget is allocated; while for high-level nodes, as the data undergoes multi-layer aggregation and processing, the privacy risk is relatively reduced, and a smaller privacy budget is allocated.

[0047] The beneficial effects of the present invention are:

[0048] Compared with the existing technology, in the present invention, at the architectural level, the multi-layer federated hierarchical learning architecture organizes the distributed participants in an orderly manner, the middle-layer aggregation nodes reduce the amount of data transmission, reduce the pressure on the central server, lay the foundation for privacy protection and parameter aggregation, and solve the problems of data silos and efficiency.

[0049] To protect privacy, differential privacy noise is added to the bottom and middle nodes, combined with dynamic privacy budget allocation to build a multi-layered protection barrier to prevent privacy leaks during model parameter transmission, ensuring the security of original data and avoiding private data leakage. At the same time, noise intensity and budget allocation are flexibly adjusted according to the scenario and training stage to balance privacy and performance.

[0050] In terms of training optimization, local model training maximizes data value and avoids privacy risks; weighted average aggregation takes into account differences in data volume and improves model stability; central server updates and global model distribution promote collaborative evolution; regular performance evaluation and hyperparameter adjustment ensure continuous model optimization, improve accuracy and generalization capabilities, and meet the needs of large-scale distributed scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic flow diagram of the present invention; DETAILED DESCRIPTION

[0052] In the description of the present invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limiting the present invention.

[0053] refer to Figure 1 In order to solve the problems existing in the background technology, this application proposes the following technical solution: a federated hierarchical learning method based on differential privacy, which specifically includes the following steps:

[0054] Step 1: System architecture design: building a multi-layer federated hierarchical learning architecture, including the bottom layer, middle layer, and top layer;

[0055] In step 1, we design the system architecture and build a multi-layer federated hierarchical learning architecture. The specific method is as follows:

[0056] First, suppose there are L layers in the system, with the bottom layer being l = 1, which contains n1 data owners (such as terminal devices or edge nodes);

[0057] Among them, the middle layer is l=2,…,L-1 layer, and each layer contains a different number of aggregation nodes;

[0058] Finally, the top layer is l=L layer, which has a central server and uses N i Represents the number of nodes in the lth layer. For node i in the lth layer, it receives model parameters from the nodes in the l-1th layer and uploads the aggregated model parameters to the nodes in the l+1th layer (when l=L, it uploads to the central server; when l=1, it receives local data for training).

[0059] The above technical solution is explained as follows: This step constructs a multi-layer federated hierarchical learning architecture, clarifying the number of nodes, roles, and data transmission direction at each layer. This hierarchical architecture organizes large-scale, distributed participants in an orderly manner, allowing data and model parameters to flow efficiently between different layers. On the one hand, the aggregation nodes in the middle layer play a role in data screening and preliminary processing, greatly reducing the amount of data directly transmitted to the central server, reducing the computing pressure and communication load of the central server, and improving the overall operating efficiency of the system. On the other hand, the clear hierarchical division makes the responsibilities of each node clear, facilitating management and maintenance, and is particularly suitable for complex scenarios where data is dispersed across different regions and institutions. At the same time, this architecture lays the foundation for subsequent hierarchical privacy protection and parameter aggregation strategies, ensuring the orderly development of the entire federated learning process, and making the system scalable and stable in large-scale applications.

[0060] Step 2: Local model training. At the bottom layer (l=1), each data owner i has a local dataset D i , the number of samples is |D i |, data owner i based on local dataset D i , use the stochastic gradient descent optimization algorithm to train the local model where θ i are model parameters;

[0061] In step 2, in the tth round of training, the local model parameter update formula is:

[0062]

[0063] Where η is the learning rate, Indicates that based on the local dataset D i Calculated loss function About the model parameter θ i The gradient of , (x, y) is the sample and label in the dataset.

[0064] The above technical solution is explained as follows: local model training is performed at the lowest level, and each data owner uses a local dataset and optimization algorithm to update model parameters. This operation fully utilizes the value of local data and avoids the privacy risks and communication costs associated with centralized data transmission. Each data owner can train the model based on the characteristics of their own data, making the model more closely aligned with the local data distribution and enhancing the model's personalization and adaptability. Moreover, the local training process is independent and unaffected by data and training status from other nodes, ensuring the autonomy and stability of the training. In this way, local data privacy is protected while decentralized data can be leveraged to jointly improve global model performance, achieving a balance between data privacy protection and model training effectiveness, effectively solving the data silo and privacy leakage issues of traditional centralized learning.

[0065] Step 3: Adding differential privacy noise (bottom-level nodes). To protect the data privacy of the bottom-level nodes, differential privacy noise is added to them before uploading the model parameters.

[0066] In step 3, according to the Laplace mechanism, the model parameter update amount of the bottom node i is Add noise z i , get the update amount after adding noise

[0067] in,

[0068]

[0069] Among them, z i It follows the Laplace distribution The noise vector is Δf, which is the sensitivity (here refers to the maximum change in the amount of model parameter update). ∈ is the privacy budget, which controls the strength of privacy protection. The smaller ∈ is, the higher the privacy protection is, but it may have a greater impact on model performance.

[0070] The above technical solution is explained as follows: Differential privacy noise is added to the lowest-level nodes before uploading model parameters, leveraging the Laplace mechanism to provide strict data privacy protection. This step effectively prevents the leakage of sensitive data at the lowest level through model parameters. Even if a malicious actor obtains the uploaded model parameters, the noise interference prevents accurate inference of the true information of the original data. This protection method provides strong protection for the data privacy of the lowest-level data owners while ensuring data availability. Furthermore, by properly setting the privacy budget, the noise intensity can be flexibly adjusted to meet different privacy protection requirements, balancing privacy protection with model performance. For scenarios with extremely high privacy requirements, the privacy budget can be appropriately reduced to enhance privacy protection; for scenarios with high model performance requirements, the privacy budget can be appropriately increased to reduce the impact of noise on the model to a certain extent, ensuring that the system meets privacy protection and performance requirements in different application scenarios.

[0071] Step 4: Parameter aggregation from the bottom layer to the middle layer, the middle layer node j (l = 2) receives the model parameter update amount after adding noise from multiple nodes in the bottom layer Perform aggregation operations;

[0072] In step 4, the weighted average aggregation method is used, and the updated model parameters after aggregation are Δθ j for:

[0073]

[0074] Among them, S j represents the set of nodes at the bottom layer that upload parameters to the middle layer node j. This weighted average aggregation takes into account the differences in the amount of data from different nodes, making the aggregation result more representative.

[0075] The above technical solution is explained as follows: a weighted average aggregation method is used to aggregate the model parameters of the bottom-level nodes to the middle-level nodes. This method fully considers the differences in the amount of data at different bottom-level nodes, so that the aggregation results can more accurately reflect the overall data characteristics. Nodes with larger data volumes have a greater weight in the aggregation, and their contribution to the global model can also be reasonably reflected, avoiding model deviations caused by uneven data volume. At the same time, the aggregation operation at the middle layer reduces the amount of data transmitted upward and reduces communication overhead. Compared to directly uploading all bottom-level node parameters to the central server, after aggregation at the middle layer, only the aggregation results are transmitted upward, which greatly reduces network transmission pressure and improves data transmission efficiency. Moreover, this aggregation method helps to smooth out the differences in data distribution at the bottom-level nodes, so that the model can still effectively learn common knowledge even when the data characteristics of different nodes are inconsistent, thereby improving the stability and generalization ability of the global model.

[0076] Step 5: Adding differential privacy noise to the middle layer (middle-layer nodes). To further protect the privacy of the middle-layer nodes and prevent the privacy information of the bottom-layer nodes from being leaked through the aggregation results, differential privacy noise is added again before the middle-layer nodes upload the aggregated parameter updates to the higher layers.

[0077] In step 5, the aggregation parameter update Δθ for the middle layer node j is j , add noise z j , get the update amount after adding noise

[0078] in,

[0079] The z here j Also follows the Laplace distribution

[0080] Where Δf′ is the sensitivity of the intermediate layer aggregate parameter update amount, and ∈′ is the privacy budget of the intermediate layer. Generally, according to the allocation strategy of the overall privacy budget of the system, ∈′ is reasonably set to balance the privacy protection of each layer and the model performance.

[0081] The above technical solution is explained as follows: Adding noise again before the intermediate nodes upload the aggregated parameter updates further strengthens privacy protection. This not only protects the privacy information of the intermediate nodes themselves, but also prevents the privacy information of the bottom-level nodes from being leaked through the intermediate-level aggregation results. Even if the intermediate nodes are attacked, it is difficult for the attacker to obtain valuable private information from the aggregated parameters after adding noise. In addition, by rationally allocating the privacy budget of the intermediate layers, the noise intensity can be flexibly adjusted according to the importance and privacy risk of the intermediate nodes, maximizing privacy protection without affecting model performance. This dual noise addition mechanism forms a multi-layered privacy protection barrier at the bottom and intermediate layers, effectively resisting privacy attacks from different levels and providing a more reliable guarantee for the privacy security of the entire federated learning system.

[0082] Step 6: Aggregate and transmit parameters between the middle layers;

[0083] In step 6, steps 4 and 5 are repeated for the intermediate layers (l = 3, ..., L-1). The nodes in the previous layer transmit the aggregated parameter updates after adding noise to the nodes in the next layer. The nodes in the next layer then perform weighted average aggregation, add noise again, and transmit the update upward. This layered parameter aggregation and noise addition method gradually transmits information from the lowest nodes upward while protecting the privacy of nodes at each layer.

[0084] The above technical solution is explained as follows: repeated parameter aggregation and noise addition operations between intermediate layers allow data to be continuously screened, aggregated, and protected during upward transmission. This layered processing method continuously reduces the amount of data transmitted and reduces communication overhead. As data is passed upward layer by layer, the amount of data ultimately reaching the central server is significantly reduced, effectively alleviating network congestion and the processing pressure on the central server. At the same time, the aggregation and noise addition of each layer further enhances privacy protection. Even if a security vulnerability occurs in a node at a certain layer, the attacker can only obtain local information after the noise is added, making it difficult to restore the complete private data. Moreover, this layered operation enables model training to better adapt to the characteristics of data at different levels, gradually refining more global and representative model parameters, which helps to improve the convergence speed and performance of the global model, enabling the system to operate efficiently and stably in a large-scale, distributed environment.

[0085] Step 7: Update the central server model;

[0086] In step 7: the specific method for updating the central server model is as follows:

[0087] First, the central server (l = L) receives the aggregate parameter update after adding noise from the top intermediate nodes Update the global model parameters θ:

[0088]

[0089] Where T represents the top-level set of intermediate nodes that upload parameters to the central server. After multiple iterations, the global model parameters are continuously updated, and the global model gradually converges.

[0090] The above technical solution is explained as follows: The central server receives the aggregated parameter updates from the top-level intermediate nodes and updates the global model parameters. In this way, the central server can integrate the information of each bottom-level node and the intermediate-level nodes to continuously optimize the global model. As the number of training rounds increases, the global model gradually converges, better fitting the overall data distribution, and improving the model's accuracy and generalization capabilities. The centralized update operation of the central server ensures the consistency and coordination of the model, enabling each node to conduct the next round of training based on a unified global model, promoting the collaborative optimization of the model across the entire system. Moreover, due to the aggregation and noise addition operations in the intermediate layer, the data received by the central server has been processed for privacy protection, reducing the risk of privacy leakage faced by the central server and ensuring the security and reliability of the system.

[0091] Step 8: Distribute the global model;

[0092] In step 8: the specific method of global model distribution is as follows:

[0093] Update the global model parameters θ on the central server t+1 After that, the updated global model is distributed to nodes at each layer;

[0094] After receiving the global model, each layer of nodes uses it as the initial model for the next round of local model training and continues the local model training process in step 2, forming a closed-loop federated hierarchical learning process.

[0095] The above technical solution is explained as follows: The central server distributes the updated global model to nodes at each layer, forming a closed-loop federated learning process. This enables each node to promptly obtain the latest global model knowledge and incorporate it into the next round of local model training, promoting the collaborative evolution of the model. Each node trains based on the updated global model, continuously adjusting its own model parameters to better align with the overall data characteristics, thereby improving the consistency and accuracy of the model. At the same time, this closed-loop process ensures the continuous operation and optimization of the system, allowing the global model to be gradually improved through continuous iteration and adapt to dynamic changes in data. Moreover, through model distribution, each node does not need to retrain the model, reducing the consumption of computing resources and improving training efficiency. This enables the system to efficiently complete large-scale model training tasks with limited resources.

[0096] Step 9: Allocate privacy budget;

[0097] In step nine, the specific method for allocating privacy budget is as follows:

[0098] In order to balance the privacy protection and model performance of the system, a dynamic privacy budget allocation strategy is designed.

[0099] Let the total privacy budget ∈ total Assign to different layers according to their importance and privacy risk level;

[0100] Assume that the privacy budget allocated to the lth layer is ∈ l ,satisfy

[0101] The lowest-level nodes, which directly access raw data and pose a higher privacy risk, are allocated a relatively larger privacy budget. Meanwhile, higher-level nodes, as data undergoes multiple layers of aggregation and processing, face lower privacy risks and are allocated a smaller privacy budget. Furthermore, the privacy budgets of each layer are dynamically adjusted based on the number of training rounds. In the early stages of training, the privacy budget is appropriately reduced to ensure rapid model convergence. Later in training, the privacy budget is increased to enhance privacy protection.

[0102] The above technical solution is explained as follows: A dynamic privacy budget allocation strategy allocates privacy budgets based on the importance of each layer, the degree of privacy risk, and the number of training rounds. At the lowest level, due to direct access to raw data, a larger privacy budget is allocated, effectively protecting the privacy of sensitive data. At higher levels, as data is aggregated and processed, privacy risks decrease, and a smaller budget is allocated to avoid excessive noise that could impact model performance. The budget is dynamically adjusted based on the number of training rounds. Reducing the budget in the early stages of training helps the model converge faster and improves training efficiency; increasing the budget in the later stages of training strengthens privacy protection and ensures data security. This strategy achieves a dynamic balance between privacy protection and model performance, enabling the system to achieve optimal operation at different stages and in different scenarios. It meets stringent requirements for data privacy protection while ensuring good model accuracy and generalization, improving the system's practicality and adaptability.

[0103] Step 10: Model Performance Evaluation and Optimization. During training, regularly evaluate the global model's performance. Use the test dataset to calculate performance metrics such as accuracy, recall, and mean squared error. Based on the performance evaluation results, adjust and optimize hyperparameters such as the learning rate and privacy budget allocation strategy to improve model performance and convergence speed. For example, if the model converges too slowly, increase the learning rate appropriately. If model performance is significantly impacted by privacy protection, readjust the privacy budget allocation to improve model performance while maintaining a certain level of privacy protection.

[0104] The above technical solution is explained as follows: The global model's performance is regularly evaluated, and hyperparameters are adjusted based on the evaluation results. By calculating performance metrics such as accuracy, recall, and mean squared error, model issues, such as slow convergence and low accuracy, can be promptly identified. To address these issues, hyperparameters such as the learning rate and privacy budget allocation strategy can be adjusted to effectively optimize model performance. For example, increasing the learning rate can accelerate model convergence, while readjusting the privacy budget allocation can improve model accuracy while ensuring privacy. This dynamic performance evaluation and optimization mechanism enables the model to continuously adapt to data changes and application requirements, continuously improving its performance. This prevents performance degradation or stagnation during training, ensuring that the model maintains high quality and reliability, providing more accurate and effective services for real-world applications and enhancing the system's competitiveness and practicality in real-world scenarios.

[0105] In the present invention, at the architectural level, a multi-layer federated hierarchical learning architecture organizes distributed participants in an orderly manner, and the middle-layer aggregation nodes reduce the amount of data transmission, reduce the pressure on the central server, lay the foundation for privacy protection and parameter aggregation, and solve the problems of data silos and efficiency.

[0106] In this paper, differential privacy noise is added to the bottom and middle nodes to protect privacy. Combined with dynamic privacy budget allocation, this creates a multi-layered protection barrier to prevent privacy leaks during model parameter transmission, ensuring the security of original data and minimizing the leakage of private data. Furthermore, noise intensity and budget allocation are flexibly adjusted based on the scenario and training stage to balance privacy and performance.

[0107] In the present invention, in terms of training optimization, local model training maximizes data value and avoids privacy risks; weighted average aggregation takes into account differences in data volume and improves model stability; central server updates and global model distribution promote collaborative evolution; regular performance evaluation and hyperparameter adjustment ensure continuous optimization of the model, improve accuracy and generalization capabilities, and meet the needs of large-scale distributed scenarios.

[0108] In summary, the present invention effectively solves the problems existing in traditional federated hierarchical learning through the deep integration of multi-layer architecture design and differential privacy technology, bringing significant improvements to data-driven machine learning applications.

[0109] From a privacy protection perspective, model training and parameter transmission in traditional federated hierarchical learning pose privacy leakage risks. However, this solution effectively prevents sensitive data from being leaked through model parameters by adding differential privacy noise before uploading parameters to the lowest-level nodes, combined with the Laplace mechanism. Intermediate-level nodes then add noise again, building a multi-layered privacy protection barrier to further prevent the leakage of privacy information from the lowest-level nodes through aggregation results. Even if the intermediate-level nodes are attacked, attackers will find it difficult to obtain valuable private information. Furthermore, a dynamic privacy budget allocation strategy rationally allocates privacy budgets based on the importance of each layer, the degree of privacy risk, and the number of training rounds. A larger budget is allocated to the lowest level to protect the privacy of raw data, while a smaller budget is allocated to higher levels to avoid excessively impacting model performance. Flexible adjustments are made at different stages of training, achieving a dynamic balance between privacy protection and model performance, comprehensively guaranteeing data privacy and security.

[0110] In terms of resolving data silos and improving training efficiency, this solution adopts a multi-layer federated hierarchical learning architecture to organize large-scale, distributed participants in an orderly manner. Local model training is performed at the bottom layer to fully utilize the value of local data and avoid the privacy risks and communication costs of centralized data transmission. Each data owner trains a model based on their own data, enhancing the model's personalization and adaptability. The aggregation nodes in the middle layer perform weighted average aggregation on the parameters of the bottom-level nodes, taking into account the difference in data volume to make the aggregation results more representative, while significantly reducing the amount of data transmitted upward, reducing communication overhead, and improving data transmission efficiency. The central server receives the aggregated parameters, updates the global model, and distributes them, forming a closed-loop process to promote model co-evolution. Each node is trained based on a unified global model to improve model consistency and accuracy, effectively resolving the data silo problem, improving training efficiency and model performance, and is suitable for large-scale, distributed scenarios.

[0111] In terms of data compliance, a strict differential privacy protection mechanism is used to ensure that the use of data in the entire federated learning process complies with privacy regulations, avoiding compliance issues caused by data leakage. This provides reliable technical support for data-driven applications, allowing enterprises and institutions to utilize the value of data without worrying about data compliance risks, thereby promoting the healthy development of machine learning technology in a big data environment.

[0112] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A federated hierarchical learning method based on differential privacy, characterized by: The specific steps include: Step 1: System architecture design: building a multi-layer federated hierarchical learning architecture, including the bottom layer, middle layer, and top layer; Step 2: Local model training. At the bottom layer, each data owner i has a local dataset D i , the number of samples is |D i |, data owner i based on local dataset D i , use the stochastic gradient descent optimization algorithm to train the local model Step 3: Add differential privacy noise. Before uploading the model parameters to the bottom-level node, add differential privacy noise to it. Step 4: Parameter aggregation from the bottom layer to the middle layer. The middle layer node j receives the model parameter update amount after adding noise from multiple nodes in the bottom layer and performs aggregation operation. Step 5: Add differential privacy noise to the middle layer. Before the middle layer node uploads the aggregated parameter update to the higher layer, differential privacy noise is added again. Step 6: Aggregate and transmit parameters between the middle layers; Step 7: Update the central server model; Step 8: Perform global model distribution; Step 9: Allocate privacy budget; Step 10: Model performance evaluation and optimization. During the training process, the performance of the global model is evaluated regularly.

2. A federated hierarchical learning method based on differential privacy according to claim 1, characterized in that: In step 1, the system architecture is designed. The specific method for building a multi-layer federated hierarchical learning architecture is as follows: First, suppose there are L layers in the system, with the bottom layer being l = 1, containing n1 data owners; Among them, the middle layer is l=2,…,L-1 layer, and each layer contains a different number of aggregation nodes; Finally, the top layer is l=L layer, which has a central server and uses N i Represents the number of nodes in layer l. For node i in layer l, it receives model parameters from nodes in layer l-1 and uploads the aggregated model parameters to nodes in layer l+1.

3. A federated hierarchical learning method based on differential privacy according to claim 2, characterized in that: In step 2, in the tth round of training, the local model parameter update formula is: Among them, θ i is the model parameter, η is the learning rate, Represents the loss function calculated based on the local dataset Dx About the model parameter θ i The gradient of , (x, y) is the sample and label in the dataset.

4. A federated hierarchical learning method based on differential privacy according to claim 3, characterized in that: In step 3, according to the Laplace mechanism, the model parameter update amount of the bottom node i is Add noise z i , get the update amount after adding noise in, Among them, z i It follows the Laplace distribution The noise vector is Δf, which is the sensitivity (here refers to the maximum change in the amount of model parameter update). ∈ is the privacy budget, which controls the strength of privacy protection. The smaller ∈ is, the higher the privacy protection is, but it may have a greater impact on model performance.

5. A federated hierarchical learning method based on differential privacy according to claim 4, characterized in that: In step 4, the weighted average aggregation method is used, and the updated model parameters after aggregation are Δθ j for: Among them, S j Indicates the bottom-level node set that uploads parameters to the middle-level node j.

6. A federated hierarchical learning method based on differential privacy according to claim 5, characterized in that: In step 5, the aggregation parameter update Δθ for the middle layer node j is j , add noise z j , get the update amount after adding noise in, The z here j Also follows the Laplace distribution Where Δf′ is the sensitivity of the update amount of the intermediate layer aggregation parameters, and ∈′ is the privacy budget of the intermediate layer.

7. A federated hierarchical learning method based on differential privacy according to claim 6, characterized in that: In step 6, for the middle layers of l=3,…,L-1, the operations of steps 4 and 5 are repeated.

8. The federated hierarchical learning method based on differential privacy according to claim 7, characterized in that: In step 7: the specific method for updating the central server model is as follows: First, the central server receives the aggregated parameter updates after adding noise from the top intermediate nodes. Update the global model parameters θ: Where T represents the set of top-level intermediate nodes that upload parameters to the central server.

9. The federated hierarchical learning method based on differential privacy according to claim 8, characterized in that: In step 8: the specific method of global model distribution is as follows: Update the global model parameters θ on the central server t+1 After that, the updated global model is distributed to nodes at each layer; After receiving the global model, each layer of nodes uses it as the initial model for the next round of local model training and continues the local model training process in step 2, forming a closed-loop federated hierarchical learning process.

10. A federated hierarchical learning method based on differential privacy according to claim 9, characterized in that: In step nine, the specific method for allocating privacy budget is as follows: Let the total privacy budget ∈ total Assign to different layers according to their importance and privacy risk level; Assume that the privacy budget allocated to the lth layer is ∈ l ,satisfy For the lowest-level nodes, since they are in direct contact with the original data, the privacy risk is higher, and a relatively larger privacy budget is allocated; while for high-level nodes, as the data undergoes multi-layer aggregation and processing, the privacy risk is relatively reduced, and a smaller privacy budget is allocated.

Images