Method for protecting privacy of personalized location association based on dynamic selection

Abstract

This invention presents a personalized location association privacy protection method based on dynamic selection, addressing the limitations of personalized performance improvement and the lack of privacy protection for parameter locations. Implementation includes: initializing and generating a private key; local training with dynamic parameter location selection; client uploading shared parameters and generating anonymous identifiers; server aggregating and distributing parameters based on a hash table; and updating the local model. This invention improves the personalized performance of the model through a dynamic parameter selection mechanism, protects the privacy of personalized parameter locations through an anonymous identifier mechanism, and utilizes cuckoo hashing for efficient storage and secure aggregation. This invention does not introduce complex cryptographic calculations, achieves personalized parameter location privacy protection, and enhances personalized performance, privacy protection capabilities, and scalability. While ensuring data and parameter location privacy and security, it achieves efficient and scalable personalized federated learning. It has applications in healthcare, finance, smart terminals, and the Internet of Things.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and information security technology, and mainly involves personalized federated learning. Specifically, it is a personalized location association privacy protection method based on dynamic selection, which can be applied to credit risk control, medical imaging, digital advertising and other fields. Background Technology

[0002] With the rapid development of information technology and network infrastructure, human society has fully entered the era of big data. Data, hailed as the "oil" of the digital economy, has become a key element driving the advancement of artificial intelligence technology. While the scale of data is growing exponentially, its use and flow face unprecedented legal and ethical constraints. In recent years, the continuous improvement of data privacy protection regulations has posed a severe challenge to traditional centralized machine learning models, which require centralized storage and processing of user data. Federated learning, proposed by Google in 2016, is a distributed machine learning paradigm where "data does not leave the local machine." By training models locally on each client and sharing only model parameters or gradients, it achieves multi-party collaborative modeling, mitigating the risk of data privacy leakage of model parameters to some extent. However, in practical applications, the data distribution of different clients often differs significantly in quantity or type, exhibiting non-independent and identically distributed characteristics. In this scenario, the single global model trained by traditional federated learning is difficult to simultaneously meet the performance requirements of all clients. Therefore, personalized federated learning, by introducing personalized parameters or personalized modeling mechanisms, enables models to better adapt to the local data distribution of each client, and has gradually become an important research direction in federated learning.

[0003] Currently, parameter-decoupled personalized federated learning is one of the most representative technical approaches. This type of method divides model parameters into shared parameters and personalized parameters, aggregating only the shared parameters globally. This balances collaborative learning and personalization needs to some extent, while reducing the number of uploaded parameters to lower communication overhead. Existing parameter-decoupled personalized federated learning typically falls into three categories: 1) using model parameters closer to the input (i.e., feature extraction layers) as personalized parameters; 2) setting classifier layers closer to the output as personalized parameters; and 3) selecting some sub-network parameters from the entire model structure as personalized parameters. The first two methods usually determine the location of personalized parameters based on network layers or predefined rules, maintaining this division throughout the training process. However, the importance of parameters in a model is not entirely determined by the network layer structure. Although the importance varies between different layers, parameters within the same layer may also have significantly different contributions to the model. Therefore, layer-based personalized parameter selection may incorrectly aggregate key parameters that should be retained locally, thus weakening the personalization effect. In contrast, the third type of fine-grained personalized parameter selection strategy can more fully leverage the advantages of parameter decoupling. However, existing subnetwork selection methods are usually static, and the model parameters of the subnetwork are not changed in multiple rounds of training. Once some parameters are selected as personalized parameters, they will be permanently retained locally and will not be uploaded or aggregated.

[0004] To improve the performance of personalized federated learning, Tamirisa et al. proposed the FedSelect method at the IEEE / CVF Conference on Computer Vision and Pattern Recognition in 2024. This method identifies parameters that change significantly across adjacent training epochs as personalized parameters. Once selected, these parameters are permanently stored locally and no longer participate in subsequent global uploads and aggregations. As training epochs progress, the number of personalized parameters gradually increases, and the local model gradually detaches from the global aggregation process, eventually evolving into a fully localized model. Therefore, the excessive personalization strategy in this method weakens the knowledge-sharing capabilities between clients, thus contradicting the original intention of federated learning to improve overall generalization performance through collaborative training.

[0005] Meanwhile, to prevent servers from inferring client privacy information through model updates during the federated learning process, some federated learning research has introduced techniques such as privacy computing to protect model parameter values. However, in personalized federated learning scenarios, the introduction of mechanisms such as parameter decoupling and differential aggregation means that privacy risks not only originate from the parameter values ​​themselves but may also be implicit in the location of personalized parameters. Existing research shows that clients with similar data distributions often exhibit consistency in model representation and parameter update patterns. This means that even if parameter values ​​are encrypted or perturbed, the selected personalized parameter locations may still be used by the server to infer similarity relationships between clients, leading to the leakage of sensitive information.

[0006] In summary, although existing research has made some progress in improving personalized performance and protecting parameter numerical privacy, federated learning cannot achieve better model training in non-independent and identically distributed scenarios. The model parameter selection strategy still suffers from the problem of limiting personalized performance development due to static partitioning. At the same time, personalized federated learning methods usually lack privacy protection measures for personalized parameter positions and do not protect the structural privacy risks inherent in parameter positions. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a personalized location association privacy protection method based on dynamic selection that breaks the association of client identity information, aiming to improve the security and personalization performance of personalized federated learning.

[0008] This invention is a personalized location association privacy protection method based on dynamic selection, comprising the following steps:

[0009] The federal learning system is provided by an authoritative agency. ,server and It consists of several clients. Responsible for generating the common parameters required by the federated learning system. The client is responsible for aggregating model parameters, while the client is responsible for training the model locally and uploading the model parameters.

[0010] (1) Initialize the federated learning system, generating private keys for anonymous identifiers during the process: initialization includes authoritative institutions ,server as well as A federated learning system for individual clients; authoritative institutions Initialize and distribute the common parameters required by the system, including three independent hash functions for the cuckoo hash map. and two sheets of size Model parameter table and location mask table ; Responsible for initializing global model parameters and location mask and compare it with the global initial iteration round. and the global maximum number of iterations Distribute to all clients; Client The local image dataset held is Furthermore, it independently generates private keys for anonymous identifiers locally. ;

[0011] (2) Each client performs local training, dynamically selecting different personalized parameter positions in each round: in the first round In the round, the client Based on local data Local training is performed to obtain local model parameters. Subsequently, the client Calculate the first Model parameters Changes in this round of training And calculate the overall importance score. The client selects the top-ranked entries based on their overall importance score. The model parameters are used as the personalized parameters for this round. and in the location mask The corresponding position will be marked as "1", and the remaining model parameters will be used as shared parameters. The position is still recorded as "0"; subsequently, the client uses the mask... Update model parameters; model parameters are divided into: In each round, each client needs to dynamically select different personalized parameter positions based on the overall importance score.

[0012] (3) Each client stores and uploads parameters based on the Cuckoo Hash and generates anonymous identifiers: Each client stores all personalized parameters for the current round. Set to zero to generate the processed model parameters. Use a mask complement Indicate shared parameters Position; for each shared parameter The hash function using the Cuckoo Hash algorithm Locate its storage location and store the value in the model parameter table. get The corresponding position markers are stored in the position mask table. get ;Then, Calculate the first The anonymous identifier for the wheel is Subsequently, the client will populate the two hash tables. and Along with anonymous identifiers Package them together and upload them to the server;

[0013] (4) The server aggregates globally shared parameters based on a hash table and distributes the results: The server aggregates the parameters uploaded by all clients using a hash table. Perform bucket-level summation to generate a globally shared parameter table. Next, the server processes the mask hash table uploaded by all clients. Perform the same bucket-by-bucket summation operation to generate the aggregate mask table. The server performs bucket-by-bucket division on these two newly generated hash tables, forming an aggregated globally shared parameter hash table. Subsequently, the server sends the table to all clients.

[0014] (5) Each client integrates local personalized parameters to update the local model: After receiving the aggregation results, the client updates the local model. Then, use the same set of hash functions as in the upload phase. The query logic retrieves the updated shared parameter values ​​from the global hash table. Combined with local personalized parameter values To update the local model parameters, as shown below. Then return to step (2) to perform the next round of federated learning training until the global maximum number of iterations is reached. This enables personalized location-associated privacy protection based on dynamic selection.

[0015] This invention addresses the limitations of personalized performance enhancement and lack of privacy protection for parameter positions in existing personalized federated learning. It achieves dynamic parameter selection and privacy protection for personalized parameter positions, and uses Cuckoo Hash to complete lightweight model parameter storage and aggregation, thereby improving the personalized performance, privacy protection capabilities, and scalability of personalized federated learning.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] (1) Create a dynamic personalized parameter selection mechanism based on historical scores to improve personalization capabilities: This invention combines the historical importance evaluation results of parameters to adaptively divide the model parameters, so that the position of personalized parameters can be dynamically adjusted with the training rounds. This avoids the excessive personalization or insufficient collaboration caused by the static decoupling strategy of the existing technology, and effectively improves the personalization effect and overall generalization ability of the model in the scenario of non-independent and identically distributed data.

[0018] (2) Enhanced privacy protection: This invention uses key-based hash message authentication code to calculate dynamic anonymous identifiers, making the parameter upload behavior of the client in different training rounds unlinkable at the identity level, weakening the direct correlation between the client identity and parameter position pattern, effectively reducing the risk of the server using model structure information to infer the client identity and data distribution, and enhancing the structural privacy protection capability in personalized federated learning scenarios.

[0019] (3) Create a structured storage and bucket-level aggregation mechanism for parameters to improve scalability: This invention achieves efficient organization and aggregation of parameters based on cuckoo hashing without introducing complex cryptographic operations. This avoids the high computation and communication overhead caused by existing technologies that rely on homomorphic encryption or multi-party secure computation, reduces the computational complexity and communication overhead of the system, and improves the scalability of the federated learning system in large-scale scenarios. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] Example 1: Data, as the fifth major factor of production, has become a key element driving the advancement of artificial intelligence technology. While the scale of data is growing, its use also faces numerous privacy risks. Federated learning, as a machine learning training paradigm that ensures data doesn't leave the local machine, reduces these risks and has become a new research direction. However, a single global model trained by federated learning struggles to adapt to the local data of all clients in non-independent and identically distributed scenarios. Personalized federated learning, by introducing personalized modeling mechanisms, has achieved some success in non-independent and identically distributed scenarios, but existing personalized federated learning schemes use static model parameter partitioning, making it difficult to adapt to dynamic changes in model parameters during training. Furthermore, the lack of privacy protection schemes for personalized parameter locations poses a risk of client privacy information leakage. Therefore, this invention explores how to propose a technical solution that can protect the location privacy of personalized parameters while dynamically selecting model parameters and improving personalized performance without increasing system overhead.

[0023] This invention is a personalized location association privacy protection method based on dynamic selection. See [link to related document]. Figure 1 , Figure 1 The flowchart of the present invention includes the following steps:

[0024] This invention is an improvement on a personalized federated learning scheme, and the federated learning system used is provided by an authoritative institution. ,server and It consists of several clients. Responsible for generating the common parameters required by the federated learning system. The client is responsible for aggregating model parameters, while the client is responsible for training the model locally and uploading the model parameters.

[0025] (1) Initialize the federated learning system, generating private keys for anonymous identifiers during the process: initialization includes authoritative institutions ,server as well as A federated learning system for individual clients. Authoritative body. Initialize and distribute the common parameters required by the system, including three independent hash functions for the cuckoo hash map. and two sheets of size Model parameter table and location mask table . Responsible for initializing global model parameters and location mask and compare it with the global initial iteration round. and the global maximum number of iterations Distribute to all clients. Client The local image dataset held is Furthermore, it independently generates private keys for anonymous identifiers locally. The private key of this invention is owned locally by the client and cannot be obtained by other clients or the server. This private key is used to generate anonymous identifiers, ensuring that the location association information of local data and parameters cannot be identified and linked by the server or other clients throughout the federated learning process, thereby enhancing the privacy protection of client data and model parameters.

[0026] (2) Each client performs local training, dynamically selecting different personalized parameter positions in each round: in the first round In the round, the client Based on local data Local training is performed to obtain local model parameters. Subsequently, the client... Calculate the first Model parameters Changes in this round of training And calculate the overall importance score. The client selects the top-ranked candidates based on their overall importance score. The model parameters are used as the personalized parameters for this round. and in the location mask The corresponding position will be marked as "1", and the remaining model parameters will be used as shared parameters. The position is still recorded as "0". Subsequently, the client uses the mask... Update model parameters; model parameters are divided into: In each training round, this invention allows each client to dynamically select different personalized parameter positions based on a comprehensive importance score. By comprehensively scoring the model parameters, different personalized parameters can be selected in each training round, achieving dynamic parameter selection and ensuring that personalized parameters can be flexibly adjusted as training rounds are updated. It also supports the global aggregation of shared parameters.

[0027] (3) Each client stores and uploads parameters based on the Cuckoo Hash and generates anonymous identifiers: Each client stores all personalized parameters for the current round. Set to zero to generate the processed model parameters. Use a mask complement Indicate shared parameters The location. For each shared parameter The hash function using the Cuckoo Hash algorithm Locate its storage location and store the value in the model parameter table. get The corresponding position markers are stored in the position mask table. get Because the Cuckoo Hash algorithm scrambles the order of the original model parameters, even if the stored results are uploaded to the server, the server cannot restore the original order of the model parameters. Therefore, this invention achieves privacy protection for personalized parameter positions. Then, Generate the first based on its own private key. The anonymous identifier for the wheel is The anonymous identifier of this invention breaks the association between client identity information and model parameter location, ensuring the anonymity and untraceability of parameter uploads. Subsequently, the client populates the two hash tables. and Along with anonymous identifiers Package them together and upload them to the server.

[0028] (4) The server aggregates globally shared parameters based on a hash table and distributes the results: The server aggregates the parameters uploaded by all clients using a hash table. Perform bucket-level summation to generate a globally shared parameter table. Next, the server processes the mask hash table uploaded by all clients. Perform the same bucket-by-bucket summation operation to generate the aggregate mask table. The server performs bucket-by-bucket division on these two newly generated hash tables, forming the aggregated globally shared parameter hash table. In this invention, the server aggregates model parameters based on a hash table, achieving correct aggregation of model parameters without exposing the original parameter locations on the client side, while also protecting the privacy of parameter locations. Subsequently, the server uses the aggregated globally shared parameter hash table... Distribute to all clients.

[0029] (5) Each client integrates local personalized parameters to update the local model: After receiving the aggregation results, the client updates the local model. Then, use the same set of hash functions as in the upload phase. The query logic retrieves the updated shared parameter values ​​from the global hash table. Combined with local personalized parameter values To update the local model parameters, as shown below. Then return to step (2) to perform the next round of federated learning training until the global maximum number of iterations is reached. This enables personalized location-associated privacy protection based on dynamic selection.

[0030] This invention is a comprehensive technical solution that addresses the limitations of personalized performance development and the lack of privacy protection for personalized parameter positions in existing personalized federated learning systems. It improves upon the static parameter selection strategies in existing personalized federated learning by proposing a dynamic parameter selection strategy based on comprehensive importance scoring. This strategy allows for flexible determination of personalized parameter positions in each training round, enhancing personalized performance. To address the lack of privacy protection for personalized parameter positions in existing systems, an anonymous identifier mechanism is designed to effectively break the association between client identity and personalized parameter positions, protecting the privacy of personalized parameter positions. Furthermore, this invention uses the Cuckoo Hash algorithm to store and upload model parameters, combined with a lightweight hash table aggregation mechanism, achieving secure aggregation and efficient distribution of shared parameters. This ensures privacy protection during the federated learning process while improving the system's scalability.

[0031] This invention presents a personalized location association privacy protection method based on dynamic selection, addressing the limitations of personalized performance improvement and lack of privacy protection for personalized parameter locations in existing personalized federated learning systems. It proposes a holistic technical solution that balances performance and privacy. This invention enhances the personalized performance of the model through a dynamic parameter selection mechanism based on comprehensive importance scoring, effectively breaks the association between client identity and personalized parameter locations through an anonymous identifier mechanism, protecting the privacy of personalized parameter locations, and utilizes cuckoo hashing to achieve efficient storage and secure aggregation of shared parameters. Without introducing complex cryptographic calculations, it achieves personalized parameter location privacy protection, improving the personalized performance, privacy protection capabilities, and scalability of the federated learning system. This invention can be applied to multi-institutional medical image analysis and personalized diagnostic model training in the healthcare field, user behavior modeling in the financial field, and personalized recommendation and model updates in smart terminal and IoT scenarios, achieving efficient and scalable personalized federated learning while ensuring data and model parameter location privacy security.

[0032] Example 2: Personalized Location-Associated Privacy Protection Method Based on Dynamic Selection. Similar to Example 1, the initialization described in step (1) of this invention... The client, the first Client Holding local image datasets .

[0033] In this embodiment, the client group in the federated learning system is initialized, and the number of clients is selected. . No. Client The local image datasets are all derived from the training set of the CIFAR-10 image classification dataset, and the number of target categories corresponding to this training set is [number missing]. It contains 60,000 images, each with a pixel size of [missing information]. There are 6,000 images for each target category.

[0034] In this embodiment, the training data is divided using a Dirichlet distribution to simulate a non-independent and identically distributed data scenario. The concentration parameter of the Dirichlet distribution is adjusted accordingly. This controls the degree of heterogeneity in data distribution between different clients, and ensures that the training samples held by each client do not overlap.

[0035] This invention, through the aforementioned client initialization and data partitioning methods, can construct non-independent, identically distributed data scenarios with varying degrees of heterogeneity under controllable conditions, closely reflecting the reality of uneven client data distribution in federated learning applications. Employing a Dirichlet distribution for data partitioning allows for adjustable data heterogeneity among different clients, facilitating a systematic evaluation of the applicability and effectiveness of the proposed dynamic personalized parameter selection and location-related privacy protection mechanism under complex data distribution conditions. Simultaneously, it ensures that the training samples of each client do not overlap, providing a reasonable and standardized foundation for subsequent personalized parameter selection, anonymous identifier generation, and parameter aggregation processes.

[0036] Example 3: Personalized location association privacy protection method based on dynamic selection. Same as Examples 1-2, the authoritative institution mentioned in step (1) of this invention. Three independent hash functions initialized for the cuckoo hash map and two sheets of size Model parameter table and location mask table .

[0037] In this embodiment, Three independent hash functions initialized From the hash function family Three independent hash functions are randomly selected from the data, and their mapping forms are as follows:

[0038] ;

[0039] in, This represents the number of buckets in the hash table and the number of parameters to be inserted. Must meet ,set up Based on the hash function described above, Initialize two images of size All-zero hash table: model parameter table Used for storing shared parameter values ​​uploaded by the client; location mask table This is used to store the location information corresponding to the parameters later.

[0040] Through authoritative institutions This embodiment uses a unified initialization of multiple independent hash functions to provide a definite and consistent mapping rule for the subsequent parameter storage and aggregation process based on cuckoo hashing. This avoids parameter conflicts or aggregation errors caused by inconsistent hash functions between clients. Furthermore, the use of multiple hash functions to construct the cuckoo hash structure allows for multiple candidate mappings of model parameters in the hash table, effectively reducing the probability of hash collisions and improving the stability and success rate of parameter storage and retrieval. In addition, by setting separate model parameter tables and location mask tables, parameter values ​​and their location information are structurally separated. This provides a foundation for subsequent shared parameter aggregation and client model updates without exposing the original model structure, enhancing the overall security of the solution in terms of parameter location-related privacy protection.

[0041] Example 4: Personalized location association privacy protection method based on dynamic selection. Same as Examples 1-3, the server mentioned in step (1) of this invention. Initialize global model parameters and location mask and compare it with the global initial iteration round. and the global maximum number of iterations Distribute to all clients.

[0042] In this embodiment, model parameters A ResNet-18 residual neural network architecture is employed. The ResNet-18 model comprises sequentially connected initial convolutional layers, multiple residual feature extraction modules, and a classification module. The initial convolutional layers extract image features. The residual feature extraction modules consist of multiple cascaded residual blocks, and feature information is transferred across layers via residual connections to improve the training stability and feature representation capability of the deep network. The classification module includes a global average pooling layer and a fully connected layer. The global average pooling layer compresses the spatial dimension of the residual features, and the fully connected layer outputs a one-dimensional vector of length 10, corresponding to the prediction results for the 10 target categories in the CIFAR-10 dataset. (Location mask) Is related to model parameters A network structure of the same dimension, initialized with all zeros. Global maximum number of iterations. It is distributed by the server to all clients.

[0043] Example 5: Personalized Location Association Privacy Protection Method Based on Dynamic Selection. Similar to Examples 1-4, the locally generated private key mentioned in step (1) of this invention... The generation method is as follows:

[0044] ;

[0045] in The private key, a cryptographically secure pseudo-random number generator, is at least 256 bits long. In this invention, each client independently invokes the cryptographically secure pseudo-random number generator to generate its private key. The private keys generated by different clients are independent of each other and statistically unpredictable. Through the above private key generation steps, it is ensured that each client independently generates a highly random and unpredictable private key locally, avoiding the risk of identity association caused by centralized key generation or reuse. This enhances the decoupling capability between the client's anonymous identifier and parameter positions, providing a reliable security foundation for subsequent privacy protection mechanisms based on personalized parameter positions.

[0046] Example 6: Personalized location association privacy protection method based on dynamic selection. Same as Examples 1-5, the client described in step (2) of this invention. Based on local data Local training is performed to obtain local model parameters. The process is as follows:

[0047] Each client Image dataset As a local model The input is forward-propagated to obtain the predicted label set for all images. Then, the cross-entropy loss function is used to... and Calculate the loss value Then through For weight parameters Find the partial derivative For the local model Update the parameters to obtain the first... Model parameters for the second training session , , The calculation formulas are as follows:

[0048] ;

[0049] ;

[0050] in, It is the total number of categories. It is the predicted label set of the training sample set. It is the first In the true label vector of the nth sample The value of each element, The model is for the first The sample is predicted to belong to the first... The probability of a class This is the learning rate.

[0051] Example 7: Personalized Location-Based Privacy Protection Method Based on Dynamic Selection. Similar to Examples 1-6, in step (2) of this invention, different personalized parameter locations are dynamically selected in each round. Calculate the overall importance score of the model parameters The calculation steps include:

[0052] (2a) Calculate the parameter change: The client expands the model parameter vector by dimension, and denotes the change as follows: The model parameters are: , its first Wheel change It is determined by the difference in parameters before and after training, and the calculation formula is:

[0053] ;

[0054] (2b) Normalization: The client normalizes the changes in all model parameters to eliminate the impact of differences in parameter scales on importance assessment, thus obtaining the relative importance score of each model parameter in the current round:

[0055] ;

[0056] in The total number of model parameters is used to normalize the overall changes. This relative importance score reflects the relative contribution of each parameter in the current training round.

[0057] (2c) Calculate the overall importance score: The client weights and fuses the relative importance score obtained in the current round with the parameter importance score from the previous round to obtain a smoothed overall importance score:

[0058] ;

[0059] in For the parameter importance score of the previous round, This is a smoothing factor used to control the fusion ratio between historical information and information from the current round.

[0060] This invention, through the aforementioned comprehensive importance scoring mechanism, simultaneously incorporates parameter change information from the current training round and importance assessment results from historical rounds during the parameter importance evaluation process. This effectively reduces the sensitivity of parameter importance to random fluctuations in a single training round, improves the stability and reliability of the importance assessment results, supports each client to dynamically select different personalized parameter positions in different training rounds, avoids the personalization rigidity problem caused by static parameter division, and further enhances the unpredictability of personalized parameter position patterns, reducing the risk of privacy leakage caused by parameter position patterns.

[0061] Example 8: Personalized Location-Associated Privacy Protection Method Based on Dynamic Selection. Similar to Examples 1-7, the client selection method in step (2) of this invention selects the top-ranked clients based on comprehensive importance scores. The model parameters are used as the personalized parameters for this round. In this invention, because it is necessary to consider the generalization ability of shared parameters in learning, The value should not be too high, and the range should be set to... In this example, the settings are as follows: The client determines the mask based on the data. Update model parameters, specifically shared parameters. and personalized parameters The update process is as follows:

[0062] ;

[0063] in, Indicates shared parameters The first in One parameter, Indicates personalized parameters The first in One parameter, Indicates mask The first in Each parameter position.

[0064] This invention, through the aforementioned model parameter update steps based on parameter location masks, can clearly distinguish between shared and personalized parameters locally on the client side using unified rules. This allows different types of parameters to play different roles during training, uploading, and aggregation, avoiding confusion between shared and personalized parameters. At the same time, by restricting personalized parameters to updates locally on the client side, it effectively prevents them from participating in global aggregation. This not only ensures the model's personalization capabilities but also provides a clear and stable parameter partitioning foundation for subsequent privacy protection mechanisms based on parameter location.

[0065] Example 9: Personalized Location-Associated Privacy Protection Method Based on Dynamic Selection. Similar to Examples 1-8, in step (3) of this invention, each client stores and uploads parameters based on the Cuckoo Hash algorithm, using the hash function of the Cuckoo Hash algorithm. Find each shared parameter The storage location is determined, and the values ​​are stored in the model parameter table. get The corresponding position markers are stored in the position mask table. get To achieve synchronous storage of parameter numerical information and positional structure information, the steps include:

[0066] (3a) Calculate the preferred position of the cuckoo hash and insert it: based on the parameter number Calculate the preferred location:

[0067] ;

[0068] If the hash bucket at that location If empty, will Insert index is hash bucket In the middle, the corresponding position mask Insert index is hash bucket middle;

[0069] (3b) Find a new cuckoo hash insertion position: If the position is not empty, evict the existing element at that position. and Insert parameters and And based on the kicked-out element index Calculate new position ;

[0070] (3c) Inserting a new element: If If the value is empty, the evicted element is stored. If the value is not empty, the element is evicted according to the above process, and a new position is searched for for the evicted element until all parameters are successfully stored.

[0071] This invention, through the aforementioned parameter storage and uploading steps based on cuckoo hashing, can structurally organize the numerical information of shared parameters using hash buckets as the basic unit. This allows the server to complete parameter reception and aggregation solely based on hash buckets, without needing to know the index position of the parameters in the original model, thus mitigating the risk of directly exposing parameter position information. Simultaneously, cuckoo hashing can achieve efficient insertion through a limited number of evictions, avoiding the introduction of complex cryptographic calculations while ensuring storage success rate. This effectively reduces the computational and communication overhead between the client and server, improving the scalability and engineering deployability of the federated learning system in large-scale scenarios.

[0072] Example 10: Personalized Location Association Privacy Protection Method Based on Dynamic Selection. Similar to Examples 1-9, the client-side calculation of the first location in step (3) of this invention... The anonymous identifier for the wheel is The calculation formula is:

[0073] ;

[0074] ;

[0075] in, This represents a key-based hash message authentication code. Indicates the first Client of , Indicates the current round number. This refers to the underlying cryptographic hash function, such as SHA-256. This is the private key held locally by the client. To input the message, and These represent the internal and external padding constants defined in the HMAC specification, respectively. This indicates a bitwise XOR operation. This indicates the concatenation of strings or byte sequences.

[0076] This invention binds anonymous identifiers to training rounds through the aforementioned anonymous identifier generation steps, relying on a client-side local private key. This allows the anonymous identifier to dynamically update with the round number across different training rounds, ensuring that the anonymous identifiers generated by the same client in different training rounds are distinct. This prevents direct association between the client's parameter upload behavior across different training rounds at the identity level. Even if the same client participates in model training across multiple rounds, its corresponding anonymous identifier will dynamically change with each round, preventing the server from tracking client parameter position patterns across rounds based on long-term identifiers. This mechanism effectively weakens the correlation between client identity and parameter position structure, achieving privacy protection for personalized parameter position identity associations without relying on complex encryption aggregation protocols, thus enhancing privacy protection capabilities.

[0077] Example 11: Personalized Location Association Privacy Protection Method Based on Dynamic Selection. Similar to Examples 1-10, the server described in step (4) of this invention performs global shared parameter aggregation based on a hash table. The implementation steps are as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] Among them, a hash table of parameters uploaded by all clients and mask hash table Perform bucket-level summation to obtain the aggregation parameter table. and aggregate mask table Then, perform bucket-by-bucket division to form the aggregated global shared parameter hash table. .

[0082] This invention employs a hash bucket-based parameter aggregation method, enabling the server to globally aggregate shared parameters without needing to recover or align the original parameter indices. This effectively avoids the server directly sensing the original position indices of model parameters. Simultaneously, by combining position masks with bucket-level aggregation, only model parameters marked as shared parameters participate in global updates, achieving decoupled aggregation of shared and personalized parameters. This aggregation method does not rely on complex cryptographic protocols such as homomorphic encryption or secure multi-party computation. While ensuring the correctness of federated learning aggregation, it significantly reduces server-side computational complexity and communication overhead, improving the scalability and engineering deployability of the method in large-scale federated learning scenarios. Furthermore, because parameters are aggregated in hash bucket form, the server cannot deduce the client's model parameter position selection through the aggregation process, further reducing the privacy risks associated with model parameter information leakage.

[0083] Example 12: A personalized location-based privacy protection method based on dynamic selection. Similar to Examples 1-11, step (5) of this invention involves each client integrating local personalized parameters to update the local model, including the following steps:

[0084] (5a) Locating the storage location of the cuckoo hash: According to Shared parameters stored in Location The corresponding parameters are calculated using the cuckoo hash function. Storage location in the updated global shared parameter hash table ;

[0085] (5b) Update model parameters: retrieve the parameter value at this location. Fill in the local model At the corresponding position:

[0086] ;

[0087] Locations not marked as shared parameters retain their local individual parameters, resulting in updated model parameters. Then return to step (2) to proceed to the next round of federated learning training.

[0088] This invention, through the aforementioned local model parameter update method, enables the client to accurately map shared parameters obtained from global aggregation back to their corresponding positions in the local model, achieving effective collaboration of shared knowledge and continuous retention of personalized model capabilities. Since the location mapping of shared parameters is completed only locally on the client, the server remains unaware of the true location structure of the model parameters in the original model throughout the entire training and aggregation process. This further blocks any potential correlation between parameter locations and client data distribution, enhancing structural privacy protection in the overall federated learning process. Simultaneously, this fusion update process does not require the introduction of additional encryption / decryption or complex communication protocols; it relies solely on hash location mapping to complete the model update, effectively reducing computational overhead and implementation complexity on the client side and improving the method's deployability in real-world scenarios.

[0089] In summary, this invention is a privacy-preserving method for personalized location association based on dynamic selection. It solves the problems of limited personalized performance development and lack of privacy protection for personalized parameter locations in existing personalized federated learning. The implementation includes: initializing the federated learning system and generating a private key for anonymous identifiers during the process; each client performing local training and dynamically selecting different personalized parameter locations in each round; each client storing and uploading parameters based on a cuckoo hash and generating anonymous identifiers; the server aggregating globally shared parameters based on a hash table and distributing the results; and each client integrating local personalized parameters to update its local model.

[0090] This invention proposes a dynamic parameter selection strategy based on comprehensive importance scoring, enabling flexible selection of personalized parameter positions in each training round. Simultaneously, an anonymous identifier mechanism is designed to effectively break the association between client identity and personalized parameter positions, protecting the privacy of personalized parameter positions. Furthermore, this invention uses the Cuckoo Hash algorithm to store and upload model parameters, combined with a lightweight hash table aggregation mechanism, achieving secure aggregation and efficient distribution of shared parameters. This invention improves the personalization performance of personalized federated learning through dynamic parameter selection, enhances privacy protection through the anonymous identifier mechanism, and improves scalability through lightweight Cuckoo Hash table aggregation.

[0091] This invention is applicable to personalized federated learning applications with high requirements for data privacy and personalized model performance, especially in collaborative modeling scenarios where data is not centralized or is semi-trusted. Specifically, this invention can be applied to multi-institutional medical image analysis and personalized diagnostic model training in the healthcare field, user behavior modeling and risk assessment in the financial field, personalized recommendation and model updates in smart terminal and IoT scenarios, as well as artificial intelligence systems involving multi-entity data collaboration such as smart cities and connected vehicles, achieving efficient and scalable personalized federated learning while ensuring data and model parameter privacy and security.

Claims

1. A personalized location association privacy protection method based on dynamic selection, characterized in that, The steps include the following: The federal learning system is provided by an authoritative agency. ,server and It consists of several clients. Responsible for generating the common parameters required by the federated learning system. The client is responsible for aggregating model parameters, while the client is responsible for training the model locally and uploading the model parameters. (1) Initialize the federated learning system, generating private keys for anonymous identifiers during the process: initialization includes authoritative institutions ,server as well as A federated learning system for individual clients; authoritative institutions Initialize and distribute the common parameters required by the system, including three independent hash functions for the cuckoo hash map. and two sheets of size Model parameter table and location mask table ; Responsible for initializing global model parameters and location mask and compare it with the global initial iteration round. and the global maximum number of iterations Distribute to all clients; Client The local image dataset held is Furthermore, it independently generates private keys for anonymous identifiers locally. ; (2) Each client performs local training, dynamically selecting different personalized parameter positions in each round: in the first round In the round, the client Based on local data Local training is performed to obtain local model parameters. Subsequently, the client Calculate the first Model parameters Changes in this round of training And calculate the overall importance score. ; The client selects the top-ranked clients based on their overall importance score. The model parameters are used as the personalized parameters for this round. and in the location mask The corresponding position will be marked as "1", and the remaining model parameters will be used as shared parameters. The position is still recorded as "0"; Subsequently, the client uses the mask. Update model parameters; model parameters are divided into: ; In each round, each client needs to dynamically select different personalized parameter positions based on the overall importance score; (3) Each client stores and uploads parameters based on the Cuckoo Hash and generates anonymous identifiers: Each client stores all personalized parameters for the current round. Set to zero to generate the processed model parameters. Use a mask complement Indicate shared parameters Position; for each shared parameter The hash function using the Cuckoo Hash algorithm Locate its storage location and store the value in the model parameter table. get The corresponding position markers are stored in the position mask table. get ;Then, Calculate the first The anonymous identifier for the wheel is ; The client then populates the two hash tables. and Along with anonymous identifiers Package them together and upload them to the server; (4) The server aggregates globally shared parameters based on a hash table and distributes the results: The server aggregates the parameters uploaded by all clients using a hash table. Perform bucket-level summation to generate a globally shared parameter table. Next, the server processes the mask hash table uploaded by all clients. Perform the same bucket-by-bucket summation operation to generate the aggregate mask table. The server performs bucket-by-bucket division on these two newly generated hash tables, forming an aggregated globally shared parameter hash table: ； The server then sends the table to all clients; (5) Each client integrates local personalized parameters to update the local model: After receiving the aggregation results, the client updates the local model. Then, use the same set of hash functions as in the upload phase. The query logic retrieves the updated shared parameter values ​​from the global hash table. Combined with local personalized parameter values To update the local model parameters, as shown below. Then return to step (2) to perform the next round of federated learning training until the global maximum number of iterations is reached. This enables personalized location-associated privacy protection based on dynamic selection.

2. The personalized location association privacy protection method based on dynamic selection according to claim 1, characterized in that, The locally generated private key mentioned in step (1) The generation method is as follows: ； in A cryptographically secure pseudo-random number generator.

3. The personalized location association privacy protection method based on dynamic selection according to claim 1, characterized in that, In step (2), each round of dynamic selection of different personalized parameter positions is performed on the client side. Calculate the overall importance score of the model parameters The calculation steps include: (2a) Calculate the parameter change: Expand the model parameter vector along the dimension, and denote the change in the parameter vector as follows: The model parameters are: , its first Wheel change The calculation formula is: ； (2b) Normalization: Normalize the changes to obtain a relative importance score: ； in This represents the total number of model parameters. (2c) Calculate the overall importance score: Combine historical importance information with the current round's importance assessment results to obtain a smoothed overall importance score: ； in For the parameter importance score of the previous round, As a smoothing factor; in each round of training, each client needs to score the importance of all parameters, allowing for dynamic selection of different personalized parameter positions.

4. The personalized location association privacy protection method based on dynamic selection according to claim 1, characterized in that, The client described in step (2) uses the mask. Update model parameters, specifically shared parameters. and personalized parameters The update process is as follows: ； in, Indicates shared parameters The first in One parameter, Indicates personalized parameters The first in One parameter, Indicates mask The first in Each parameter position.

5. The personalized location association privacy protection method based on dynamic selection according to claim 1, characterized in that, In step (3), each client stores and uploads parameters based on the Cuckoo Hash algorithm, using the hash function of the Cuckoo Hash algorithm. Find each shared parameter The storage location is determined, and the values ​​are stored in the model parameter table. get The corresponding position markers are stored in the position mask table. get The implementation steps include: (3a) Calculate the preferred position of the cuckoo hash and insert it: based on the parameter number Calculate the preferred location: ； If the hash bucket at that location If empty, will Insert index is hash bucket In the middle, the corresponding position mask Insert index is hash bucket middle; (3b) Find a new cuckoo hash insertion position: If the position is not empty, evict the existing element at that position. and Insert parameters and And based on the kicked-out element index Calculate new position ; (3c) Inserting a new element: If If the value is empty, the evicted element is stored. If the value is not empty, the element is evicted according to the above process, and a new position is searched for for the evicted element until all parameters are successfully stored.

6. The personalized location association privacy protection method based on dynamic selection according to claim 1, characterized in that, The client described in step (3) Calculate the first The anonymous identifier for the wheel is The formula for calculating anonymous identifiers is: ； ； in, This represents a key-based hash message authentication code. Indicates the first Client of , Indicates the current round number. This refers to the underlying cryptographic hash function, such as SHA-256. This is the private key held locally by the client. To input the message, and These represent the internal and external padding constants defined in the HMAC specification, respectively. This indicates a bitwise XOR operation. This indicates the concatenation of strings or byte sequences.

7. The personalized location association privacy protection method based on dynamic selection according to claim 1, characterized in that, The server described in step (4) performs global shared parameter aggregation based on a hash table. The implementation steps are as follows: ； ； ； Among them, a hash table of parameters uploaded by all clients and mask hash table Perform bucket-level summation to obtain the aggregation parameter table. and aggregate mask table Then, perform bucket-by-bucket division to form the aggregated global shared parameter hash table: 。 8. The personalized location association privacy protection method based on dynamic selection according to claim 1, characterized in that, Step (5) involves each client integrating local personalized parameters to update the local model, including the following steps: (5a) Locating the storage location of the cuckoo hash: According to Shared parameters stored in Location The corresponding parameters are calculated using the cuckoo hash function. Storage location in the updated global shared parameter hash table ; (5b) Update model parameters: retrieve the parameter value at this location. Fill in the local model At the corresponding position: ； Generate updated model parameters Then return to step (2) to proceed to the next round of federated learning training.

Images