Federal learning privacy protection and communication optimization method and system

Abstract

The invention relates to the technical field of machine learning, in particular to a federated learning privacy protection and communication optimization method and system, and the method comprises the steps: obtaining a local training data set of a client, and carrying out the preprocessing based on the gradient compression and gradient rarefaction technology; constructing an adaptive scrambling mechanism, and generating a training sequence scrambling mechanism based on the dynamic estimation value of the privacy disclosure risk; performing distribution probability density estimation of a training sequence scrambling mechanism, performing subspace division on the training sample set, and performing probability density estimation and aggregation based on locality sensitive hashing on each subspace; secure aggregation of encryption parameters is realized based on a homomorphic encryption technology and a locality sensitive hashing technology; each client updates a local model based on the aggregation parameters and implements incremental updating and differential transmission, and realizes refined privacy protection through an adaptive scrambling mechanism; the communication data volume is reduced through the gradient compression and sparse cooperation technology, and collaborative optimization of privacy protection and communication efficiency is achieved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for privacy protection and communication optimization in federated learning, which solves the key problems of data privacy protection and communication efficiency in the federated learning process. Background Technology

[0002] With increasing awareness of data privacy, federated learning, as a distributed machine learning framework for protecting data privacy, has received widespread attention. Federated learning allows multiple participants to collaboratively train models without sharing the original data, effectively solving the problems of data silos and privacy protection. However, existing federated learning technologies still face two key challenges: firstly, traditional federated learning frameworks lack effective privacy protection mechanisms and are vulnerable to privacy threats such as gradient backpropagation attacks and model inversion attacks; secondly, the frequent transmission of large amounts of model parameters between servers and clients leads to high communication overhead and low training efficiency, especially in environments with limited network bandwidth.

[0003] In existing technologies, methods such as differential privacy and homomorphic encryption are generally used for privacy protection. However, these methods often employ globally fixed privacy protection strategies, failing to provide fine-grained protection based on the sensitivity and importance of different parameters, easily leading to over-protection or under-protection. Meanwhile, in terms of communication optimization, while existing techniques such as gradient compression and gradient quantization can reduce communication volume to some extent, important information may be lost during compression, affecting model convergence and accuracy. Furthermore, privacy protection and communication optimization in existing technologies are often separate, lacking effective collaborative optimization mechanisms, making it difficult to achieve a good balance between ensuring privacy and improving communication efficiency.

[0004] Therefore, there is an urgent need for a federated learning optimization method that can simultaneously address the issues of privacy protection and communication efficiency, thereby achieving synergistic optimization of privacy protection and communication efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for privacy protection and communication optimization in federated learning, aiming to solve the problems of insufficient privacy protection, low communication efficiency, and difficulty in coordinating privacy protection and communication optimization in existing federated learning.

[0006] This invention proposes a method for privacy protection and communication optimization in federated learning, including:

[0007] Obtain the local training dataset on the client side;

[0008] The local training dataset is preprocessed to obtain preprocessed training data, wherein the preprocessing includes:

[0009] The local training data is processed using gradient compression and gradient sparsification techniques to obtain compressed and sparsified gradient parameters.

[0010] An adaptive scrambling mechanism is constructed to scramble the training sequences, wherein the construction of the adaptive scrambling mechanism includes:

[0011] Based on a dynamic estimate of privacy leakage risk, a scrambling mechanism is generated for training sequences.

[0012] The dynamic estimate of the privacy leakage risk is used as the input parameter generated by the training sequence scrambling mechanism;

[0013] Estimating the probability density distribution of the training sequence scrambling mechanism includes:

[0014] The training sample set is divided into subspaces to form multiple training sample subsets;

[0015] Perform probability density estimation based on locality-sensitive hashing on each subset of training samples;

[0016] The probability density estimates on each subset are aggregated to obtain the global probability density estimate.

[0017] Based on homomorphic encryption and locality-sensitive hashing techniques, secure aggregation of encryption parameters is achieved, including:

[0018] Send the encrypted gradient parameters to the aggregator;

[0019] During the aggregation phase, locality-sensitive hashing homomorphic encryption is used to globally aggregate the parameters sent by each client;

[0020] The aggregation results are decrypted and distributed to each client node;

[0021] Each client node updates its local model based on the received aggregation parameters and feeds the update results back to the federated learning server.

[0022] Preferably, the gradient compression and gradient sparsity techniques include:

[0023] Based on a predefined data sparsity, the gradients obtained during training are sparsified while retaining important gradient information. The determination of important gradient information is based on the L2 norm of the weight vector.

[0024] The non-zero gradients after sparsification are compressed using a scaling factor to further reduce the amount of data transmitted.

[0025] The gradient is compressed into an integer representation of a specified number of bits through the quantization process, reducing communication costs.

[0026] Preferably, the scrambling mechanism for generating training sequences based on the dynamic estimate of privacy leakage risk includes:

[0027] Calculate the sensitivity of model parameters and identify high-risk parameters;

[0028] Evaluate the degree of influence of each parameter on the inference result;

[0029] A dynamic risk index is generated based on the results of sensitivity analysis and impact assessment.

[0030] An adaptive noise matrix is ​​generated based on the dynamic risk index to ensure that high-risk parameters are better protected.

[0031] The scrambling strategy is dynamically adjusted as the training progresses, with increased protection strength in the early training phase and gradually reduced during model convergence.

[0032] Preferably, the subspace partitioning of the training sample set includes:

[0033] Based on the data source, an initial subset is formed;

[0034] A secondary partitioning based on feature similarity is performed to obtain the final training sample subset;

[0035] For samples that fall on the boundaries of a subspace, a soft assignment strategy is adopted, allowing them to belong to multiple subspaces simultaneously.

[0036] A special allocation algorithm is used to ensure that the number of samples in each subspace is relatively balanced, thus avoiding data skew.

[0037] Preferably, performing probability density estimation based on locality-sensitive hashing on each subset of training samples includes:

[0038] Based on the feature distribution of training samples within the subspace, an initial probability density function is established;

[0039] The most suitable kernel function is adaptively selected based on data characteristics;

[0040] The kernel function bandwidth is dynamically adjusted according to the data distribution to achieve a balance between smoothness and accuracy;

[0041] The density function is continuously optimized as the training process progresses through an incremental update mechanism, avoiding redundant calculations.

[0042] Preferably, the aggregation of probability density estimates on each subset includes:

[0043] Weights are assigned to the estimation results of each subspace based on the number and representativeness of the subspace samples;

[0044] Special treatment is applied to the subspace boundary region to ensure a smooth transition;

[0045] Identify and process estimation results that deviate significantly from the main distribution;

[0046] The aggregation results are normalized to ensure that the sum of probabilities is 1.

[0047] Preferably, the secure aggregation of encryption parameters based on homomorphic encryption and locality-sensitive hashing includes:

[0048] Locality-sensitive hashing is used to guide the encryption process, so that similar data use the same encryption parameters.

[0049] A layered encryption strategy is adopted for model parameters with different sensitivities;

[0050] Parallel encryption processing is achieved by mapping model parameters to different hash buckets using locality-sensitive hash functions.

[0051] By leveraging the properties of homomorphic encryption, we can directly perform addition operations on the encrypted parameters to achieve parameter aggregation in the encrypted state.

[0052] Preferably, after each client node updates its local model, the method further includes:

[0053] Perform incremental updates and difference transfers, including:

[0054] Record historical values ​​of model parameters and identify areas of significant change;

[0055] Only parameters that have changed relative to the previous version and their changed values ​​are transmitted;

[0056] Predictions based on parameter change trends can further reduce the amount of content transmitted.

[0057] Implement communication congestion control and scheduling optimization, including:

[0058] Sensing the network status and computing power of each client;

[0059] Set transmission priorities for different types of parameters;

[0060] Bandwidth allocation is dynamically adjusted based on the training phase and parameter importance.

[0061] Preferably, the method also includes client contribution assessment:

[0062] The locality-sensitive hashing algorithm is used to map high-dimensional gradient data to multiple hash buckets;

[0063] Calculate the gradient statistics for each bucket;

[0064] Evaluate each client's contribution to each hash bucket;

[0065] The overall contribution index of the client is formed by combining the contribution evaluation results of each hash bucket;

[0066] Adjust the client's weight in model updates based on contribution metrics.

[0067] The Federalized Learning Privacy Protection and Communication Optimization System includes:

[0068] The client-side data processing module is used to preprocess the local training dataset and generate compressed and sparsified gradient parameters based on gradient compression and gradient sparsification techniques.

[0069] The server-side coordination and management module is used to build an adaptive scrambling mechanism, generate a training sequence scrambling mechanism based on dynamic estimates of privacy leakage risks, and coordinate the entire federated learning process.

[0070] The probability density estimation module is used to divide the training sample set into subspaces, perform probability density estimation based on locality-sensitive hashing in each subspace, and aggregate the estimation results of each subset to form a global probability density estimate.

[0071] The secure aggregation module is used to securely aggregate encrypted parameters based on homomorphic encryption and locality-sensitive hashing techniques, and then decrypts the aggregation results and distributes them to each client.

[0072] The model update module enables each client to update its local model based on the received aggregate parameters, and implements incremental updates and differential transmission to optimize communication efficiency.

[0073] The server-side coordination and management module, probability density estimation module, and security aggregation module work together to ensure data privacy and security while achieving efficient utilization of communication resources.

[0074] This invention innovatively combines Locality Sensitive Hashing (LSH) with adaptive probability density estimation to construct a refined privacy protection mechanism that adaptively adjusts the protection strength based on the sensitivity and importance of parameters. Simultaneously, it employs gradient compression and sparsity optimization techniques to significantly reduce communication overhead. Furthermore, this invention designs a secure aggregation mechanism that integrates homomorphic encryption and LSH, achieving efficient aggregation of encryption parameters.

[0075] The present invention has the following beneficial effects:

[0076] 1. Improved the precision and efficiency of privacy protection. Through an adaptive scrambling mechanism, the protection strength is dynamically adjusted based on parameter sensitivity, avoiding over-protection or under-protection issues caused by a globally uniform protection strategy. Under the same level of privacy protection, the model accuracy is improved by approximately 18%.

[0077] 2. Significantly reduced communication overhead. The combined gradient compression and sparsification technique retains key gradient information while reducing the amount of communication data by 75% to 85%, greatly improving training efficiency.

[0078] 3. Enhanced security and efficiency of the aggregation process. Based on a secure aggregation mechanism using homomorphic encryption and locality-sensitive hashing, aggregation computation efficiency is improved by 3-5 times while ensuring data security.

[0079] 4. Achieves synergistic optimization of privacy protection and communication efficiency. This invention integrates privacy protection and communication optimization into a complete solution that performs excellently in practical applications, making it particularly suitable for fields with high requirements for data privacy and communication efficiency, such as healthcare, financial services, and smart cities. Attached Figure Description

[0080] Figure 1 A flowchart of a federated learning privacy protection and communication optimization method provided in an embodiment of the present invention;

[0081] Figure 2 This is a flowchart illustrating the adaptive scrambling mechanism generation process in an embodiment of the present invention.

[0082] Figure 3 This is a flowchart of probability density estimation based on locality-sensitive hashing in an embodiment of the present invention;

[0083] Figure 4 This is a flowchart of the secure aggregation mechanism in an embodiment of the present invention;

[0084] Figure 5 This is a structural diagram of a federated learning privacy protection and communication optimization system provided in an embodiment of the present invention. Detailed Implementation

[0085] Please refer to Figures 1-5 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the embodiments described in the present invention are merely exemplary and do not constitute a limitation on the scope of protection of the present invention.

[0086] See Figure 1 The federated learning privacy protection and communication optimization method provided by this invention includes the following steps:

[0087] First, the client obtains the local training dataset. In practical applications, this data may come from hospital patient records, bank financial transactions, user behavior of smart devices, etc. This data usually contains sensitive information and requires strict privacy protection.

[0088] Next, the local training dataset is preprocessed to obtain preprocessed training data. In a preferred embodiment of the invention, the preprocessing process uses gradient compression and gradient sparsification techniques to process the local training data, obtaining compressed and sparsified gradient parameters. This step effectively reduces the amount of data transmitted subsequently, laying the foundation for communication optimization.

[0089] Then, an adaptive scrambling mechanism is constructed to scramble the training sequences. Specifically, a scrambling mechanism for the training sequences is generated based on a dynamic estimate of the privacy leakage risk; this dynamic estimate is used as the input parameter for generating the training sequence scrambling mechanism. This adaptive scrambling mechanism can provide differentiated protection for parameters with different sensitivities, significantly improving the fineness of privacy protection.

[0090] Next, probability density estimation based on the training sequence scrambling mechanism is performed. This step includes: subdividing the training sample set into multiple training sample subsets; performing probability density estimation based on locality-sensitive hashing on each training sample subset; and aggregating the probability density estimates from each subset to obtain the global probability density estimation result. This divide-and-conquer strategy improves the accuracy of probability density estimation while reducing computational complexity.

[0091] This system achieves secure aggregation of encrypted parameters based on homomorphic encryption and locality-sensitive hashing (LSH) techniques. Specifically, it involves: sending encrypted gradient parameters to the aggregator; using LSH homomorphic encryption to globally aggregate parameters sent by each client during the aggregation phase; and decrypting the aggregation result before distributing it to each client node. This innovative secure aggregation mechanism significantly improves aggregation efficiency while ensuring data security.

[0092] Finally, each client node updates its local model based on the received aggregation parameters and feeds the updated results back to the federated learning server. This step completes one round of model updates, and the system will continue to execute the above steps until the model converges or reaches the preset number of training rounds.

[0093] This invention employs a combined gradient compression and gradient sparsification technique to process model gradients, effectively reducing communication overhead. The specific implementation of this combined technique is as follows:

[0094] In one embodiment of the present invention, the gradients obtained during training are first sparsified based on a predefined data sparsity, retaining important gradient information. The determination of important gradient information is based on the L2 norm of the weight vector. Specifically, let the gradient matrix be... The weight vector is ,in Indicates the dimension index of the weight vector. Representing feature dimension, This represents the number of weight vectors. Calculate the L2 norm of each weight vector. and compare it with a preset threshold. Comparison. When If the gradient is zero, retain the gradient of that dimension; otherwise, set the gradient of that dimension to zero.

[0095] in: This is the gradient matrix, containing the gradient values ​​of all parameters; For the first dimensional weight vector; Indicates the first The L2 norm of the dimensional weight vector is calculated as follows: ,in Indicates the first The dimensional weight vector The value of each element; A preset threshold is used to determine the importance of the weight vector.

[0096] Preferably, the predefined data sparsity is generally set between 0.1 and 0.3, and can be adjusted according to the specific application scenario and data characteristics. For example, for image data, a lower sparsity (such as 0.1) may be needed to retain more feature information; while for text data, a higher sparsity (such as 0.3) may be suitable. In practical applications, the data sparsity can also be dynamically adjusted according to the training phase, using a lower sparsity to retain more information in the early stages of training, and gradually increasing the sparsity as training progresses.

[0097] The non-zero gradient after sparsification is compressed using a scaling factor to further reduce data transmission. Let the sparsified gradient be... The scaling factor is The compressed gradient The calculation is as follows:

[0098] ,

[0099] in: This is the sparsed gradient matrix, where some elements have been set to zero; This is a scaling factor used to further compress the size of the gradient values; This is the compressed gradient matrix.

[0100] The value is usually between 0.1 and 0.5. The selection of values ​​needs to balance compression ratio and information retention rate. The smaller the value, the higher the compression ratio, but more information may be lost. The larger the value, the more complete the information is preserved, but the compression effect is weaker. In a preferred embodiment of the present invention, A value of 0.3 strikes a good balance in most application scenarios.

[0101] Quantization compresses gradients into an integer representation of a specified number of bits, further reducing communication costs. Specifically, floating-point gradients can be quantized into 8-bit integers, significantly reducing data storage space. The quantization process can be represented as:

[0102] ,

[0103] in: This is the quantized gradient matrix, with integer elements; This is a rounding function that converts a real number to the nearest integer. and They are respectively The minimum and maximum values ​​in; The number of quantization bits (usually 8) determines the quantization precision; It is the maximum representation value calculated based on the number of bits b.

[0104] By employing the aforementioned gradient compression and sparsification synergistic techniques, this invention significantly reduces the amount of communication data while preserving key gradient information. Tests show that communication overhead can be reduced by 75% to 85%, providing strong support for efficient training in federated learning.

[0105] This invention innovatively designs an adaptive scrambling mechanism based on privacy leakage risks, achieving refined privacy protection by dynamically adjusting the scrambling strength. See also... Figure 2 The specific implementation of the adaptive scrambling mechanism is as follows:

[0106] First, the sensitivity of the model parameters is calculated to identify high-risk parameters. In one embodiment of the invention, parameter sensitivity can be determined by calculating the degree of influence of the parameters on the prediction results. Specifically, for parameters... Its sensitivity It can be represented as:

[0107] ,

[0108] in: For parameters Sensitivity; This is the loss function for the model; This indicates the loss function with respect to the parameters. The partial derivatives; This indicates the operation of taking the absolute value. The higher the sensitivity, the greater the impact of this parameter on the model's prediction results, and the higher the risk of privacy leakage.

[0109] Next, assess the impact of each parameter on the inference results. This step can be achieved through perturbation analysis, which involves adding small perturbations to the parameters and observing the changes in the model output. (Influence level) It can be represented as:

[0110] ,

[0111] in: For parameters The degree of impact; The number of samples; Let J be the feature vector of the j-th sample; This is the model inference function, which outputs the prediction results. A complete set of model parameters; Indicates that the parameter The set of parameters after being set to zero; This indicates the absolute value operation. The formula calculates the result over all samples after removing the parameter. The model outputs the average absolute value of the change.

[0112] Based on the results of sensitivity analysis and impact assessment, a dynamic risk index is generated. The dynamic risk index is calculated by comprehensively considering factors such as parameter sensitivity, degree of influence, and the current training round.

[0113] ,

[0114] in: This is a dynamic risk index, representing the degree of privacy leakage risk of the parameters; and These are normalized vectors representing sensitivity and impact, respectively, with the original values ​​normalized to the [0,1] interval; This is the current training round; The preset total number of training rounds; It is an exponentially decaying term that gradually decreases as training progresses; , and Let be the weighting coefficient, satisfying It is used to adjust the importance of various factors.

[0115] In a preferred embodiment of the present invention , , These parameter values ​​have been verified through extensive experiments and demonstrate good performance in various application scenarios. (Exponential term) The introduction of scrambling gradually weakens the scrambling intensity as training progresses, which helps improve model convergence.

[0116] An adaptive noise matrix is ​​generated based on the dynamic risk index to ensure stronger protection for high-risk parameters. (Noise matrix) The generation can be represented as:

[0117] ,

[0118] in: This is the noise matrix, with the same dimensions as the parameter matrix; This refers to the dynamic risk index calculated above; Based on the base noise intensity, control the overall noise level; This indicates that the mean is 0 and the covariance matrix is ​​the identity matrix. The random noise value is generated by applying a multivariate Gaussian distribution. In this way, parameters with higher risk indices will have stronger noise added, thus achieving more stringent protection.

[0119] The scrambling strategy is dynamically adjusted as the training progresses, increasing the protection strength in the early training phase and gradually decreasing it during model convergence. This dynamic adjustment strategy is based on two observations: firstly, the model parameters change significantly in the early training phase, leading to a higher risk of privacy leakage; secondly, as the model converges, parameter changes decrease, allowing for a more appropriate reduction in protection strength to improve model performance. Specifically, the base noise intensity... Adjustments can be made according to the training progress:

[0120] ,

[0121] in: For the first Noise intensity during round training; The initial noise intensity represents the baseline noise level at the start of training. The attenuation coefficient controls the rate at which the noise intensity decreases over time; This is the current training round; This is the preset total number of training rounds.

[0122] Preferably, Usually set to Typically, a value between 0.5 and 0.8 is used. These parameters can be adjusted based on specific application scenarios and privacy protection requirements. Higher values... and lower Offers stronger privacy protection, but may affect model performance; lower and higher This is beneficial for model performance, but may reduce the effectiveness of privacy protection.

[0123] Through the aforementioned adaptive scrambling mechanism, this invention achieves refined privacy protection, preserving useful information to the greatest extent while ensuring privacy security and improving model performance. Experiments show that, compared with traditional fixed-noise methods, this scheme improves model accuracy by approximately 18% at the same level of privacy protection.

[0124] See Figure 3 This invention employs an innovative subspace partitioning and probability density estimation method, which improves the accuracy and efficiency of density estimation. The specific implementation is as follows:

[0125] First, the training sample set is subspaced to form multiple training sample subsets. In one embodiment of the present invention, the subspace partitioning includes a preliminary partition based on the data source to form an initial subset; and a secondary partition based on feature similarity to obtain the final training sample subset.

[0126] Preferably, the initial segmentation based on data source can be based on client ID or data type. For example, in a medical scenario, it can be segmented by hospital, department, or disease type; in a financial scenario, it can be segmented by institution, business type, or customer group. This initial segmentation fully considers the natural attributes and organizational structure of the data.

[0127] Secondary partitioning based on feature similarity utilizes clustering algorithms to group similar samples in the feature space into a single group. Specifically, clustering algorithms such as K-means and DBSCAN can be used. Taking K-means as an example, the number of subspaces K is first determined (usually 5 to 10), then the Euclidean distance or cosine similarity between samples is calculated, and iterative optimization is performed until convergence. Each cluster after clustering is considered a subspace.

[0128] For samples falling within the boundaries of subspaces, this invention employs a soft assignment strategy, allowing them to belong to multiple subspaces simultaneously. The soft assignment strategy calculates the degree of belonging based on the distance of the sample to the center of each subspace:

[0129] ,

[0130] in: Indicates sample pair of subspaces The degree of belonging, with a value range of For the sample To subspace Distance from the center; Temperature is a parameter that controls the smoothness of the affiliation distribution. This represents the total number of subspaces. This represents an exponential function. When... When the value exceeds a preset threshold (e.g., 0.2), the sample is considered invalid. Belonging to subspace .

[0131] A special allocation algorithm ensures a relatively balanced number of samples in each subspace, avoiding data skew. Specifically, minimum and maximum sample number thresholds can be set. When the number of samples in a subspace is too small, its boundary can be appropriately expanded; when the number of samples is too large, the subspace can be further subdivided. This balancing strategy helps improve the reliability of subsequent probability density estimation.

[0132] Probability density estimation based on locality-sensitive hashing is performed on each subset of training samples. First, an initial probability density function is established based on the feature distribution of the training samples within the subspace. This step can employ the kernel density estimation (KDE) method.

[0133] ,

[0134] in: Let be the estimated probability density function, and let represent the eigenvector. The probability density of occurrence; The number of samples within the subspace; For kernel functions, such as Gaussian kernel, Epanechnikov kernel, etc.; This is a bandwidth parameter that controls the smoothness. For the first subspace Feature vectors of each sample; Representing the eigenvector With sample The difference.

[0135] The most suitable kernel function is adaptively selected based on data characteristics. In a preferred embodiment of the invention, different kernel functions can be selected according to the data distribution characteristics. For example, the Gaussian kernel function performs well for approximately normally distributed data; the Epanechnikov kernel function may be more suitable for bounded data. This invention implements an automatic kernel function selection mechanism, evaluating the performance of different kernel functions through cross-validation and selecting the optimal one.

[0136] The kernel function bandwidth is dynamically adjusted based on the data distribution to achieve a balance between smoothness and accuracy. (Bandwidth) The choice of bandwidth directly affects the quality of density estimation; too small a bandwidth leads to overfitting, while too large a bandwidth leads to over-smoothing. This invention uses the Silverman rule to automatically determine the initial bandwidth:

[0137] ,

[0138] in: For bandwidth parameters; The standard deviation of the sample represents the degree of dispersion of the sample distribution. It is the interquartile range, which is the third quartile minus the first quartile; The number of samples; This indicates taking the smaller value of the two parameters. Then, cross-validation is used to further optimize the bandwidth parameters.

[0139] Through an incremental update mechanism, the density function is continuously optimized as training progresses, avoiding redundant calculations. When new samples are added, it is not necessary to recalculate the entire density function; only the affected portion needs to be updated.

[0140] ,

[0141] in: Indicates the first The density estimation function of the wheel; Indicates the first The density estimation function of the wheel; For newly added samples; This represents the number of samples that have been processed so far. For kernel functions; This is the bandwidth parameter. This incremental update mechanism significantly improves computational efficiency, especially on large-scale datasets.

[0142] The probability density estimates from each subspace are aggregated to obtain the global probability density estimate. First, weights are assigned to the estimates for each subspace based on the number and representativeness of the subspace samples. weight It can be calculated as:

[0143] ,

[0144] in: For subspace The weights represent the importance of the subspace in the global density estimation; For subspace The number of samples; For subspace Representative indicators can be assessed based on sample diversity or coverage. This represents the total number of subspaces. This represents summing over all subspaces.

[0145] Special processing is applied to the boundary regions of subspaces to ensure a smooth transition. For samples belonging to multiple subspaces, their probability density can be calculated using a weighted average.

[0146] ,

[0147] in: For the sample The global probability density estimate; For the sample pair of subspaces degree of belonging; For subspace The probability density estimation function on The value at; This represents the total number of subspaces. This represents summing over all subspaces.

[0148] Identify and process estimation results that significantly deviate from the main distribution. This can be achieved by setting a threshold. ,when At that time, the subspace is considered If the estimation results are abnormal, their weight should be appropriately reduced or they should be excluded directly.

[0149] in Representing subspace The absolute value of the difference between the estimated value and the global estimated value; This is a preset threshold, typically set to 0.3 to 0.5; This is the global estimate.

[0150] Finally, the aggregation results are normalized to ensure that the sum of probabilities is 1:

[0151] ,

[0152] in: This is the normalized probability density function; The probability density function after aggregation; This represents the integral of the probability density function over the entire feature space, i.e., the sum of probabilities.

[0153] By employing the aforementioned subspace partitioning and probability density estimation methods, this invention improves estimation accuracy while reducing computational complexity. Tests show that compared to direct global estimation, the computational complexity is reduced from O(n²) to O(n·log(n)), and the estimation accuracy is improved by approximately 15%.

[0154] See Figure 4 This invention designs an innovative secure aggregation mechanism that combines homomorphic encryption and locality-sensitive hashing, significantly improving aggregation efficiency while ensuring data security. The specific implementation is as follows:

[0155] First, locality-sensitive hashing (LSH) guides the encryption process, ensuring that similar data uses relevant encryption parameters. This step fully leverages the property of LSH: similar inputs map to similar hash values. Specifically, for the parameter vector... Calculate its locality-sensitive hash value Then based on Generate encryption parameters: in: For parameters The set of encryption parameters; Generate a function for the parameters, mapping hash values ​​to the encrypted parameter space; For parameters Locality-sensitive hash value; To enhance encryption security, a secure seed is used. This ensures that similar parameters will use relevant encryption parameters, facilitating efficient subsequent aggregation.

[0156] A layered encryption strategy is adopted for model parameters with different sensitivities. Based on the aforementioned privacy risk assessment results, the parameters are divided into three layers according to their sensitivity: high, medium, and low, and each layer employs a different encryption strategy. For example, for high-sensitivity parameters, more complex encryption algorithms or longer keys can be used; for low-sensitivity parameters, a lightweight encryption scheme can be used to reduce computational overhead.

[0157] Parallel encryption processing is achieved by mapping model parameters to different hash buckets using locality-sensitive hash functions. Specifically, an appropriate family of locality-sensitive hash functions is selected. For parameter vectors Calculate its hash value set And based on these hash values, The parameters are assigned to the corresponding hash buckets. Similar parameters are likely to be assigned to the same hash bucket, allowing them to be encrypted and processed together, thus improving parallel efficiency.

[0158] in: It is a family of locality-sensitive hash functions, including One hash function; For the first A hash function maps parameters to hash values; For parameters After the first The hash value after mapping by a hash function.

[0159] Preferably, the selection of a hash function family should consider the data characteristics. For example, for data in Euclidean space, a p-stable distributed hash function can be used; for set-type data, a MinHash function can be used. In one embodiment of the present invention, the number of hash functions m is typically 3 to 5, and the number of hash buckets B can be adjusted according to the parameter scale, generally being the square root of the total number of parameters.

[0160] By leveraging the properties of homomorphic encryption, addition operations are directly performed on the encrypted parameters to achieve parameter aggregation in the encrypted state. Homomorphic encryption is a special encryption technique that allows specific algebraic operations on encrypted data without prior decryption. In this invention, an additive homomorphic encryption scheme is employed, enabling... ,in Indicates the encryption function. and The parameters are in plaintext. This characteristic allows the aggregator to directly sum the encrypted parameters without knowing the original parameter values.

[0161] Implement weighted aggregation under encrypted conditions, supporting differentiated processing of contributions from different clients. Assume the client... The contribution weight is The parameters are The implementation of weighted aggregation is then:

[0162] ,

[0163] in: Represents a homomorphic encryption function; Indicates all Summing of data from multiple clients; For the client The contribution weight is usually a positive number and satisfies For the client Parameters; For the client The encrypted parameters. This weighted aggregation mechanism can adjust weights based on client data quality, quantity, or historical contributions, improving the quality of the aggregation results.

[0164] This invention controls precision loss during the encryption aggregation process through a special numerical representation method. Homomorphic encryption typically only supports integer operations, thus requiring the conversion of floating-point parameters to integer representations. This invention employs fixed-point representation, multiplying the floating-point number by a large integer (such as...). The data is then rounded to an integer. After decryption, it is divided by that integer to restore the floating-point representation. This method ensures the correctness of homomorphic encryption while maintaining computational precision.

[0165] A verification protocol is designed to ensure the correctness of the aggregation results. In a distributed environment, malicious clients may submit incorrect parameters. To prevent this, this invention implements a verification protocol based on zero-knowledge proofs. Each client must provide proof that its parameters conform to the expected range and format. The aggregator verifies the validity of these proofs and only accepts valid parameters for aggregation.

[0166] By employing the aforementioned homomorphic encryption and locality-sensitive hashing (LSH) collaborative security aggregation mechanism, this invention significantly improves aggregation efficiency while ensuring data security. Tests show that compared to traditional homomorphic encryption aggregation, computational efficiency is improved by 3-5 times, with particularly noticeable effects in large-scale parameter aggregation.

[0167] After completing the basic model update, this invention also implements an incremental update and differential transmission mechanism to further optimize communication efficiency. The specific implementation is as follows:

[0168] First, record the historical values ​​of the model parameters and identify the parts that change significantly. Let the current round be t, and the parameters be... The parameters for the previous round were Calculate the change in parameters Then set the threshold. ,when When, the parameters are considered Significant changes require transmission; otherwise, transmission can be omitted.

[0169] in: This is the current training round; Let be the model parameters for round t; These are the model parameters for the (t-1)th round; For parameter changes; This represents the change in the i-th parameter; Represents the absolute value of the change; This is a preset threshold used to determine whether parameter changes are significant.

[0170] Preferably, the threshold The threshold can be dynamically adjusted according to the training phase. In the early stages of training, when parameters change significantly, a higher threshold (e.g., 0.01) can be set; as training progresses and parameters gradually stabilize, the threshold can be gradually lowered (e.g., to 0.001). This dynamic threshold strategy achieves a better balance across different training phases.

[0171] Only parameters that have changed compared to the previous version and their changed values ​​are transmitted, significantly reducing the amount of data transmitted. The transmitted data can be represented as a series of tuples (i, ..., ...). ), where i is the parameter index, These are variable values. In practical implementations, sparse matrix representation can be used to further compress storage and transmission space.

[0172] Predictions based on parameter change trends can further reduce transmitted content. By analyzing historical parameter change patterns, the change trend for the current round can be predicted. Let the parameter changes in the past k rounds be... , , ..., Then the predicted current change can be expressed as:

[0173] ,

[0174] in: This represents the predicted parameter change in round t. This represents the actual parameter change in round ti; The weighting coefficient represents the degree of influence of different historical rounds on the prediction, satisfying the following conditions: k represents the number of historical rounds considered. This represents the summation over all historical cycles considered.

[0175] Generally, recent changes are given greater weight, and exponential decay weighting can be used to set the weights. ,in This is the attenuation coefficient, typically taken as 0.5 to 0.9. When the actual change... Compared with the predicted value The difference is less than the threshold At this time, only the difference value can be transmitted. This further reduces the amount of data transmitted.

[0176] This invention also implements communication congestion control and scheduling optimization, including sensing the network status and computing power of each client, setting transmission priorities for different types of parameters, and dynamically adjusting bandwidth allocation according to the training stage and parameter importance.

[0177] Specifically, network conditions can be assessed by periodically measuring round-trip time (RTT) and packet loss rate, and transmission strategies can be adjusted accordingly. For example, for clients with poor network conditions, the parameter update frequency can be appropriately reduced, or a more aggressive compression strategy can be adopted.

[0178] Assign transmission priorities to different types of parameters. Generally, shallow network parameters vary more and have a greater impact on model performance, so they should be given higher transmission priority; parameters in deep feature extraction parts are relatively stable and can be given lower priority. In implementation, parameters can be divided into multiple priority queues according to their importance and transmitted in priority order.

[0179] Bandwidth allocation is dynamically adjusted based on the training stage and parameter importance. In the early stages of training, model parameters change significantly, requiring more bandwidth allocation; as training progresses and parameters gradually stabilize, bandwidth allocation can be reduced. Simultaneously, important parameters should receive more bandwidth resources to ensure timely updates.

[0180] Through the aforementioned incremental update and differential transmission mechanisms, as well as communication congestion control and scheduling optimization, this invention further reduces communication overhead and improves the system's ability to adapt to complex network environments. Tests show that compared to basic gradient compression methods, the incremental update mechanism can additionally reduce communication by 20%–30%, with particularly significant effects during the model's convergence phase.

[0181] To prevent malicious node attacks and optimize participant selection, this invention designs a client contribution evaluation mechanism, which is implemented as follows:

[0182] First, the locality-sensitive hashing (LSH) algorithm is used to map high-dimensional gradient data to multiple hash buckets. A suitable LSH function h is selected to map the gradient vector G to B hash buckets. The choice of hash function should ensure that similar gradient vectors are likely to be mapped to the same hash bucket. In a preferred embodiment of the invention, an LSH function of the following form can be used:

[0183] ,

[0184] in: Gradient vector The hash value; Let be a random vector whose elements follow a standard normal distribution. ; Represents the dot product of vectors; for A random real number within a given range; The bucket width parameter controls the granularity of the hash bucket; This indicates a round-down operation.

[0185] To improve the accuracy of the mapping, a hash table can be constructed using m hash functions, and then L such hash tables can be built. The selection of parameters m and L should balance accuracy and computational cost; typically, m is 5-10 and L is 10-20.

[0186] Calculate the gradient statistics for each bucket. For hash buckets... Its statistical values ​​can include the mean, variance, and maximum value. Taking the mean as an example:

[0187] ,

[0188] in: For hash buckets The average value of the gradient vector; For hash buckets The number of gradient vectors in the middle; Indicates the hash bucket All gradient vectors Sum.

[0189] Evaluate each client's contribution to each hash bucket. In hash bucket Contribution in Through its gradient Measured by similarity to bucket statistics:

[0190] ,

[0191] in: For the client In hash bucket Contribution in; For the client The gradient vector; For hash buckets The gradient mean; Represents the dot product of vectors; Represents the gradient vector The L2 norm is calculated as follows: ,in Gradient vector The One element; Represents the gradient average The L2 norm.

[0192] It is actually cosine similarity, with values ​​between [-1, 1]. The larger the value, the higher the contribution.

[0193] The overall contribution index of client i is formed by combining the contribution evaluation results of each hash bucket. It can be represented as:

[0194] ,

[0195] in: The overall contribution metric for client i; Let m be the total number of hash buckets, m be the number of hash functions in each hash table, and L be the number of hash tables. The weight of hash bucket j is usually proportional to the number of gradient vectors in the bucket; The contribution of client i in hash bucket j; This represents the summation over all hash buckets.

[0196] The client's weight in the model update is adjusted based on the contribution metric. Let the initial weight of client i be... (This can be determined based on factors such as data volume), then the actual weight after considering contribution. It can be represented as:

[0197] ,

[0198] in: The adjusted weight for client i; Let i be the initial weight for client i; The contribution metric for client i; and These represent the minimum and maximum contribution values ​​among all clients, respectively; score items. The contribution is normalized to the [0,1] interval.

[0199] This adjustment gives higher weight to clients with higher contributions and lower weight to clients with lower contributions, but it does not completely exclude any client (the weight is at least 50% of the initial weight).

[0200] Through the aforementioned client contribution evaluation mechanism, this invention can effectively identify and reduce the impact of malicious or low-quality clients, improving the security and model quality of the federated learning system. Tests show that, even with 10% malicious clients, this mechanism can control model performance loss to within 5%, significantly better than the over 30% loss without protection.

[0201] See Figure 5 This invention also provides a federated learning privacy protection and communication optimization system, including a client-side data processing module, a server-side coordination and management module, a probability density estimation module, a security aggregation module, and a model update module. These modules work together to realize the various functions of the aforementioned method.

[0202] The client-side data processing module preprocesses the local training dataset, generating compressed and sparsed gradient parameters based on gradient compression and gradient sparsification techniques. This module is deployed on each client device participating in federated learning and is responsible for the preprocessing and initial compression of local data, laying the foundation for subsequent communication optimization.

[0203] The server-side coordination and management module is used to build an adaptive scrambling mechanism, generate training sequence scrambling mechanisms based on dynamic estimates of privacy leakage risks, and coordinate the entire federated learning process. Deployed on the central server, this module is responsible for global coordination and privacy risk assessment, and is the core control unit of the system.

[0204] The probability density estimation module is used to partition the training sample set into subspaces, perform probability density estimation based on locality-sensitive hashing in each subspace, and aggregate the estimation results of each subset to form a global probability density estimate. This module can be deployed on a server or a dedicated computing node, and is responsible for efficient and accurate probability density estimation, providing the basic data for scrambling mechanisms.

[0205] The secure aggregation module is used to securely aggregate encrypted parameters based on homomorphic encryption and locality-sensitive hashing (LSH) techniques. The aggregated results are then decrypted and distributed to each client. This module is typically deployed on a dedicated aggregation server and is responsible for the secure and efficient aggregation of parameters, making it a crucial step in ensuring data privacy.

[0206] The model update module enables each client to update its local model based on the received aggregate parameters, and performs incremental updates and difference transmissions to optimize communication efficiency. This module is deployed on each client device and is responsible for updating and optimizing the local model, ensuring continuous model improvement.

[0207] In the system of this invention, the server-side coordination and management module, probability density estimation module, and security aggregation module work together to ensure data privacy and security while achieving efficient utilization of communication resources. This modular design gives the system good scalability and adaptability, enabling it to meet the needs of different application scenarios.

[0208] The following specific embodiment illustrates in detail the application of the present invention in federated learning of medical data.

[0209] Suppose five hospitals participate in federated learning to jointly train a disease prediction model. Each hospital possesses its own patient data, but for privacy protection reasons, they cannot directly share the original data. The specific implementation steps of the federated learning privacy protection and communication optimization method of this invention are as follows:

[0210] First, each hospital acquires its local training dataset, which includes patients' basic information, examination results, and diagnostic labels. This data contains sensitive personal health information, requiring strict privacy protection.

[0211] Then, the local training dataset is preprocessed. The model is trained based on the local data, gradients are calculated, and gradient compression and sparsity techniques are applied to process the gradient parameters. In this embodiment, the data sparsity is set to 0.2, retaining approximately 20% of the important gradient information; the scaling factor... A value of 0.3 was used to compress non-zero gradients; finally, 8-bit quantization was used to further reduce the data volume. Tests showed that these parameter settings achieved a good balance in medical data, reducing the amount of communication data by approximately 80% while retaining key information.

[0212] Next, an adaptive scrambling mechanism is constructed to scramble the training sequences. First, the sensitivity of the model parameters is calculated to assess the influence of each parameter on the prediction results. In medical data, features directly related to diagnosis (such as specific biomarkers) typically have higher sensitivity. Then, based on sensitivity analysis and influence assessment, a dynamic risk index is formed, with weighting coefficients... , An adaptive noise matrix is ​​generated based on the dynamic risk index, with the base noise intensity... The initial value was set to 0.08 and gradually decreased during training. Experiments show that this adaptive scrambling mechanism can effectively protect sensitive medical data while maintaining the model's predictive accuracy.

[0213] Next, probability density estimation using a training sequence scrambling mechanism is performed. First, the training sample set is divided into seven subspaces, with an initial division based on disease type and a secondary division based on patient feature similarity. For boundary samples (such as patients with multiple diseases), a soft assignment strategy is used, with a membership threshold set to 0.2. Probability density estimation based on locality-sensitive hashing is then performed in each subspace, using a Gaussian kernel function, with the bandwidth parameter automatically determined using the Silverman rule. Finally, the density estimation results from each subspace are aggregated to form a global probability density estimate. This divide-and-conquer strategy significantly improves the efficiency and accuracy of density estimation.

[0214] Secure aggregation of encrypted parameters is achieved using homomorphic encryption and locality-sensitive hashing (LSH) techniques. Layered encryption based on parameter sensitivity is employed, with highly sensitive parameters (such as features related to rare diseases) receiving stronger encryption protection. LSH distributes parameters across 15 hash buckets for parallel encryption processing. Paillier homomorphic encryption is used to directly perform addition operations on the encrypted parameters, achieving secure aggregation. A 10⁶-fold fixed-point conversion is used to handle floating-point numbers. This secure aggregation mechanism significantly improves aggregation efficiency while protecting medical data privacy.

[0215] Finally, each hospital updates its local model based on the received aggregate parameters and feeds the updated results back to the federated learning server. Incremental updates and difference propagation are implemented simultaneously, with a threshold... The initial value was set to 0.01 and gradually reduced to 0.001 during training. Communication congestion control was dynamically adjusted based on the hospital's network conditions, prioritizing the transmission of parameters related to key disease characteristics. Tests showed that this optimization strategy further reduced communication overhead by approximately 25%.

[0216] In practical applications, this method has been successfully applied to the joint disease prediction model training of multiple hospitals. While protecting patient privacy, the model accuracy has reached more than 95% of that of single-center training, and the communication overhead has been reduced by more than 80%, proving the practical value of this invention.

[0217] The federated learning privacy protection and communication optimization method and system provided by this invention effectively solves the privacy protection and communication efficiency problems in federated learning through innovative adaptive scrambling mechanisms, subspace probability density estimation, homomorphic encryption, and locally sensitive hashing collaborative secure aggregation. Experiments and practical applications show that this solution significantly improves communication efficiency while ensuring data privacy, providing strong support for the widespread application of federated learning.

[0218] This invention is applicable to a variety of scenarios, including but not limited to healthcare, financial services, and smart cities, and is particularly suitable for applications that require both data collaboration and strict privacy protection. As data privacy regulations become increasingly stringent and the demand for distributed learning continues to grow, the technological value of this invention will become even more apparent.

[0219] The above description is merely a preferred embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural transformations made based on the inventive concept of the present invention and the contents of the specification and drawings of the present invention, or direct / indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for privacy protection and communication optimization in federated learning, characterized in that: include: Obtain the local training dataset on the client side; The local training dataset is preprocessed to obtain preprocessed training data, wherein the preprocessing includes: The local training data is processed using gradient compression and gradient sparsification techniques to obtain compressed and sparsified gradient parameters. An adaptive scrambling mechanism is constructed to scramble the training sequences, wherein the construction of the adaptive scrambling mechanism includes: Based on a dynamic estimate of privacy leakage risk, a scrambling mechanism is generated for training sequences. The dynamic estimate of the privacy leakage risk is used as the input parameter generated by the training sequence scrambling mechanism; Estimating the probability density distribution of the training sequence scrambling mechanism includes: The training sample set is divided into subspaces to form multiple training sample subsets; Perform probability density estimation based on locality-sensitive hashing on each subset of training samples; The probability density estimates on each subset are aggregated to obtain the global probability density estimate. Based on homomorphic encryption and locality-sensitive hashing techniques, secure aggregation of encryption parameters is achieved, including: Send the encrypted gradient parameters to the aggregator; During the aggregation phase, locality-sensitive hashing homomorphic encryption is used to globally aggregate the parameters sent by each client; The aggregation results are decrypted and distributed to each client node; Each client node updates its local model based on the received aggregation parameters and feeds the update results back to the federated learning server.

2. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, The gradient compression and gradient sparsity techniques include: Based on a predefined data sparsity, the gradients obtained during training are sparsified while retaining important gradient information. The determination of important gradient information is based on the L2 norm of the weight vector. The non-zero gradients after sparsification are compressed using a scaling factor to further reduce the amount of data transmitted. The gradient is compressed into an integer representation of a specified number of bits through the quantization process, reducing communication costs.

3. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, The scrambling mechanism for generating training sequences based on the dynamic estimate of privacy leakage risk includes: Calculate the sensitivity of model parameters and identify high-risk parameters; Evaluate the degree of influence of each parameter on the inference result; A dynamic risk index is generated based on the results of sensitivity analysis and impact assessment. An adaptive noise matrix is ​​generated based on the dynamic risk index to ensure that high-risk parameters are better protected. The scrambling strategy is dynamically adjusted as the training progresses, with increased protection strength in the early training phase and gradually reduced during model convergence.

4. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, The subspace partitioning of the training sample set includes: Based on the data source, an initial subset is formed; A secondary partitioning based on feature similarity is performed to obtain the final training sample subset; For samples that fall on the boundaries of a subspace, a soft assignment strategy is adopted, allowing them to belong to multiple subspaces simultaneously. A special allocation algorithm is used to ensure that the number of samples in each subspace is relatively balanced, thus avoiding data skew.

5. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, The step of performing probability density estimation based on locality-sensitive hashing on each subset of training samples includes: Based on the feature distribution of training samples within the subspace, an initial probability density function is established; The most suitable kernel function is adaptively selected based on data characteristics; The kernel function bandwidth is dynamically adjusted according to the data distribution to achieve a balance between smoothness and accuracy; The density function is continuously optimized as the training process progresses through an incremental update mechanism, avoiding redundant calculations.

6. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, The aggregation of probability density estimates on each subset includes: Weights are assigned to the estimation results of each subspace based on the number and representativeness of the subspace samples; Special treatment is applied to the subspace boundary region to ensure a smooth transition; Identify and process estimation results that deviate significantly from the main distribution; The aggregation results are normalized to ensure that the sum of probabilities is 1.

7. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, The secure aggregation of encryption parameters based on homomorphic encryption and locality-sensitive hashing includes: Locality-sensitive hashing is used to guide the encryption process, so that similar data use the same encryption parameters. A layered encryption strategy is adopted for model parameters with different sensitivities; Parallel encryption processing is achieved by mapping model parameters to different hash buckets using locality-sensitive hash functions. By leveraging the properties of homomorphic encryption, we can directly perform addition operations on the encrypted parameters to achieve parameter aggregation in the encrypted state.

8. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, After each client node updates its local model, the method further includes: Perform incremental updates and difference transfers, including: Record historical values ​​of model parameters and identify areas of significant change; Only parameters that have changed relative to the previous version and their changed values ​​are transmitted; Predictions based on parameter change trends can further reduce the amount of content transmitted. Implement communication congestion control and scheduling optimization, including: Sensing the network status and computing power of each client; Set transmission priorities for different types of parameters; Bandwidth allocation is dynamically adjusted based on the training phase and parameter importance.

9. The method for privacy protection and communication optimization in federated learning according to claim 1, characterized in that, This method also includes client contribution assessment: The locality-sensitive hashing algorithm is used to map high-dimensional gradient data to multiple hash buckets; Calculate the gradient statistics for each bucket; Evaluate each client's contribution to each hash bucket; The overall contribution index of the client is formed by combining the contribution evaluation results of each hash bucket; Adjust the client's weight in model updates based on contribution metrics.

10. A federated learning privacy protection and communication optimization system, characterized in that, include: The client-side data processing module is used to preprocess the local training dataset and generate compressed and sparsified gradient parameters based on gradient compression and gradient sparsification techniques. The server-side coordination and management module is used to build an adaptive scrambling mechanism, generate a training sequence scrambling mechanism based on dynamic estimates of privacy leakage risks, and coordinate the entire federated learning process. The probability density estimation module is used to divide the training sample set into subspaces, perform probability density estimation based on locality-sensitive hashing in each subspace, and aggregate the estimation results of each subset to form a global probability density estimate. The secure aggregation module is used to securely aggregate encrypted parameters based on homomorphic encryption and locality-sensitive hashing techniques, and then decrypts the aggregation results and distributes them to each client. The model update module enables each client to update its local model based on the received aggregate parameters, and implements incremental updates and differential transmission to optimize communication efficiency. The server-side coordination and management module, probability density estimation module, and security aggregation module work together to ensure data privacy and security while achieving efficient utilization of communication resources.

Images