Data development and utilization method and system based on federated learning framework

Abstract

This application discloses a data development and utilization method and system based on a federated learning framework. The method includes: receiving encrypted model data generated from local non-shared raw data transmitted by each federated learning participant through a secure communication link; performing homomorphic encryption on the model parameter gradient information in the encrypted model data and verifying it in conjunction with model structure metadata to obtain global encrypted model data that has undergone hierarchical aggregation encryption processing; distributing the global encrypted model data to each participant according to a fragmented transmission strategy adjusted based on the network bandwidth and computing resource capacity of the participants; guiding each participant to use the global encrypted model data to update parameters and optimize the structure of their local business model, and generating optimization suggestion information by evaluating the model update status through a local validation dataset. This method can achieve multi-party data cooperation and sharing while protecting data security and privacy, thereby improving model performance.

Description

Technical Field

[0001] This application relates to the field of computer data processing technology, and in particular to a data development and utilization method and system based on a federated learning framework. Background Technology

[0002] In the application of spicy seasonings, a large amount of data is generated across multiple stages, including production, sales, and marketing. The production stage encompasses detailed information on raw material procurement, such as the source, quality, and price of different ingredients, as well as production process parameters like temperature, pressure, and time control. The sales stage includes the diversity of sales channels, such as online e-commerce platforms, offline supermarkets, and convenience stores, and sales data for each channel. The marketing stage includes consumer feedback, such as evaluations of product taste, packaging, and price, as well as market research information, such as consumer preferences and competitor product information. This data is held by different entities, including manufacturers, distributors, and retailers.

[0003] Traditional data sharing and utilization methods often require the direct exchange of raw data. However, due to data security and privacy concerns, parties are reluctant to share data directly. In machine learning model training, a centralized training approach has been commonly used, where all data is concentrated at a central node for training. However, this method faces numerous challenges in the spicy seasoning scenario because the data characteristics and distributions of different participants vary significantly. Centralized training struggles to fully leverage the strengths of each party's data, leading to poor model performance. Furthermore, directly sharing raw data poses serious data security and privacy risks, potentially leaking company trade secrets.

[0004] In summary, traditional centralized training methods cannot adapt to the characteristics of scattered and highly variable data in the spicy seasoning scenario, resulting in limited model performance. In addition, directly sharing raw data poses serious data security and privacy risks, making it difficult for all parties to achieve effective data cooperation and model training. Summary of the Invention

[0005] This application provides a data development and utilization method and system based on a federated learning framework.

[0006] This application provides a data development and utilization method based on a federated learning framework, applied to a data development and utilization system, the method comprising:

[0007] Receive encrypted model data transmitted by each federated learning participant through a secure communication link. The encrypted model data is generated by each federated learning participant after completing model training locally based on non-shared raw data.

[0008] The model parameter gradient information in the encrypted model data is homomorphically encrypted using an asymmetric encryption algorithm, and then verified in conjunction with the model structure metadata in the encrypted model data to obtain the global model encrypted data that has completed the hierarchical aggregation encryption process.

[0009] The global model encrypted data is distributed to each federated learning participant according to a preset fragmentation transmission strategy, which is adjusted based on the network bandwidth and computing resource capacity of the federated learning participant.

[0010] The system guides each federated learning participant to use the global model encrypted data to update parameters and optimize the structure of their local business model, and uses the local validation datasets of each federated learning participant to evaluate the model update status and generate optimization suggestions.

[0011] One embodiment of this application provides a data development and utilization system, including:

[0012] A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; and when the computer program is executed by the processor, the processor enables the processor to implement any of the data development and utilization methods based on the federated learning framework described above.

[0013] One embodiment of this application provides a readable storage medium on which a program or instruction is stored, and when the program or instruction is executed by a processor, it implements the steps of the data development and utilization method based on the federated learning framework. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating a data development and utilization method based on a federated learning framework, provided as an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the basic structure of a data development and utilization system provided in an embodiment of this application.

[0017] Figure 3 This is a functional block diagram of a data development and utilization device based on a federated learning framework, provided for an embodiment of this application.

[0018] Figure 4This is a schematic diagram of a federated learning interactive environment provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] Please see Figure 1 , Figure 1 This is a flowchart of a data development and utilization method based on a federated learning framework provided in an embodiment of this application. The method can be executed by the data development and utilization system, or by the data development and utilization system and the server together. The method may include steps 110-140.

[0021] Step 110: Receive encrypted model data transmitted by each federated learning participant through a secure communication link. The encrypted model data is generated by each federated learning participant after completing model training locally based on non-shared original data.

[0022] In the application scenario of spicy seasonings, there are multiple participants in the federated learning process, including manufacturers, distributors, and retailers. Each participant possesses unique, non-shared raw data derived from their respective business operations. For example, manufacturers have information on raw material procurement, including the source and quality grade of different raw materials, as well as production process parameters such as temperature, pressure, and time; distributors possess detailed information on sales channels, such as cooperating supermarkets and e-commerce platforms, and sales data for each channel; retailers collect consumer feedback, such as evaluations of product taste and packaging, as well as market research information, such as consumer preferences and competitor product information.

[0023] Each participant in the federated learning system trains its model locally using its own non-shared raw data. Taking a manufacturer as an example, a Convolutional Neural Network (CNN) model from deep learning could be used. This model has an input layer, convolutional layers, pooling layers, and fully connected layers. The input layer receives data on raw material procurement and production process parameters; the convolutional layers extract features from the data using convolutional kernels; the pooling layers perform dimensionality reduction on the features; and the fully connected layers integrate the processed features and output the prediction result. During training, the backpropagation algorithm is used to adjust the model's parameters to minimize the error between the prediction and the actual results.

[0024] For example, a manufacturer's local model uses a convolutional neural network (CNN) to process raw material procurement and production process parameter data to predict production efficiency. The number of neurons in the input layer is determined by the data dimensions. If the raw material procurement data includes various indicators such as price and quality grade for 10 different raw materials, and the production process parameters include 5 key parameters such as temperature, pressure, and time, then the number of neurons in the input layer is 15, and the input data is a 15-dimensional vector.

[0025] The first convolutional layer uses a 3x3 kernel with 32 kernels. The stride is 1, and the padding is "same" to maintain the feature map size. The activation function is ReLU, which effectively mitigates gradient vanishing and makes model training more efficient. The second convolutional layer also uses a 3x3 kernel, but the number of kernels increases to 64 to extract more complex features. The stride and padding remain the same, and the activation function is still ReLU.

[0026] The pooling layer uses max pooling with a 2x2 pooling window and a stride of 2. This reduces the amount of data while preserving important feature information.

[0027] The first fully connected layer has 128 neurons, and the ReLU activation function is used to enhance non-linear expressiveness. The second fully connected layer serves as the output layer. Since the production efficiency is to be predicted into three levels—high, medium, and low—the number of neurons is 3, and the Softmax activation function is used to convert the output into a probability distribution for classification.

[0028] During training, the stochastic gradient descent (SGD) optimization algorithm was used, with a learning rate of 0.001 and a momentum parameter of 0.9. The cross-entropy loss function was used to measure the difference between the predicted and actual results, and the training epochs were 200.

[0029] For example, the dealer local model uses a variant of recurrent neural network (RNN), the Long Short-Term Memory (LSTM), to process sales channel and sales data to predict sales trends. The number of neurons in the input layer depends on the data dimension. If there are eight types of sales channels, such as online e-commerce platforms, offline supermarkets, and convenience stores, and the sales data includes four indicators, such as sales in different time periods, the number of neurons in the input layer is 12, and the input data is a 12-dimensional vector.

[0030] The number of neurons in the LSTM layer is set to 64. The LSTM unit has a forget gate, an input gate, and an output gate, which can effectively handle long-term dependencies in sequence data.

[0031] The fully connected layer has 1 neuron. The activation function uses the Sigmoid function to convert the output into a probability value between 0 and 1, representing the probability of an upward sales trend.

[0032] The Adam optimization algorithm was used for training, with a learning rate of 0.0005. The binary cross-entropy loss function was used, which is suitable for binary classification problems, and the training epochs were 150.

[0033] For example, a retailer's local model uses a multilayer perceptron (MLP) to process consumer feedback and market research data to analyze consumer preferences. The number of neurons in the input layer depends on the data dimensions. If consumer feedback includes six aspects such as taste, packaging, and price satisfaction, and market research data includes four indicators such as competitor product information and market demand trends, the number of neurons in the input layer is 10, and the input data is a 10-dimensional vector.

[0034] The first hidden layer has 50 neurons, and the activation function is ReLU. The second hidden layer has 30 neurons, and the activation function is also ReLU.

[0035] The number of neurons in the output layer is determined based on the number of consumer preference categories, which are divided into 5 different preference types. The output layer has 5 neurons and uses the Softmax activation function to output the probability distribution.

[0036] The Adagrad optimization algorithm was used for training, with a learning rate of 0.01. The cross-entropy loss function was used, and the training epochs were 180.

[0037] After training, each participant encrypts the model data. The encryption process uses an asymmetric encryption algorithm, such as RSA. RSA generates public and private keys based on the difficulty of factoring large prime numbers. Participants use the public key to encrypt the model data, ensuring its security during transmission. The encrypted model data is then transmitted to the data development and utilization system via a secure communication link. This secure communication link employs the SSL / TLS protocol, which uses encryption and authentication mechanisms to ensure that data is not stolen or tampered with during transmission. The data development and utilization system receives this encrypted model data and verifies its integrity and authenticity, for example, by checking the data's hash value to ensure that the data has not been modified during transmission.

[0038] Step 120: Perform homomorphic encryption operation on the model parameter gradient information in the encrypted model data using an asymmetric encryption algorithm, and verify it in conjunction with the model structure metadata in the encrypted model data to obtain the global model encrypted data that has completed the hierarchical aggregation encryption process.

[0039] After receiving encrypted model data from each federated learning participant, the data development and utilization system begins processing the model parameter gradient information. This gradient information reflects the changes in the model's parameters during training.

[0040] In detail, the step of performing homomorphic encryption on the model parameter gradient information in the encrypted model data using an asymmetric encryption algorithm, and then verifying it in conjunction with the model structure metadata in the encrypted model data, yields the globally encrypted model data that has undergone hierarchical aggregation encryption processing, including:

[0041] Step 121: Extract the model structure metadata from the encrypted model data, and parse it to obtain the model hierarchical structure descriptor and parameter dimension distribution features.

[0042] The data development system extracts model structure metadata from encrypted model data. This metadata includes information such as the model's hierarchical structure and parameter dimensions. During parsing, the system first identifies key information in the metadata, such as the number of layers, the type of each layer (e.g., convolutional layer, fully connected layer), and the number and dimensions of parameters in each layer. Analysis of this information yields the model hierarchical structure descriptor and parameter dimension distribution characteristics. The model hierarchical structure descriptor can be represented by a tree structure, where each node represents a model layer, and the connections between nodes represent the relationships between layers. The parameter dimension distribution characteristics record the parameter dimension information for each layer; for example, for a convolutional layer, it records the number, size, and number of channels of its convolutional kernels.

[0043] Step 122: Perform hierarchical partitioning of the model parameter gradient information based on the model hierarchy structure descriptor to generate a gradient information subset corresponding to the model hierarchy.

[0044] Based on the model hierarchy descriptor obtained in step 121, the model parameter gradient information is hierarchically divided. The system assigns the model parameter gradient information to the corresponding layers according to the model's hierarchical order. For example, for a neural network model with an input layer, hidden layers, and an output layer, the input layer's parameter gradient information is assigned to a subset of the input layer's gradient information, the hidden layer's parameter gradient information to a subset of the hidden layer's gradient information, and the output layer's parameter gradient information to a subset of the output layer's gradient information. Each gradient information subset contains all the parameter gradient information for that layer.

[0045] Step 123: Use the public key of the asymmetric encryption algorithm to perform homomorphic encryption operation on each gradient information subset to generate encrypted gradient information units.

[0046] Homomorphic encryption is performed on each subset of gradient information using the public key of an asymmetric encryption algorithm. Homomorphic encryption allows corresponding computations to be performed on encrypted data without decryption. Taking the Paillier homomorphic encryption algorithm as an example, this algorithm exhibits additive homomorphism, meaning that the result of adding two encrypted data is the same as the result of adding the decrypted data and then encrypting it. The system uses the public key to encrypt each subset of gradient information, generating encrypted gradient information units. During encryption, the integrity and security of each encrypted gradient information unit are ensured, for example, by adding an encryption checksum to verify whether the data has been tampered with.

[0047] Step 124: Perform correlation verification between the encrypted gradient information unit and the corresponding parameter dimension distribution features to verify the matching of parameter dimensions before and after encryption, and obtain the encrypted gradient set that has passed the verification.

[0048] The generated encrypted gradient information units are correlated with their corresponding parameter dimension distribution features. The system checks whether the parameter dimensions of the encrypted gradient information units are consistent with the parameter dimension distribution features. If they are inconsistent, it indicates that an error may have occurred during the encryption process or the data may have been tampered with. During the verification process, the parameter dimensions of each encrypted gradient information unit are checked one by one to ensure that they match the description in the parameter dimension distribution features. If the verification passes, these encrypted gradient information units are combined into a set of verified encrypted gradients.

[0049] Step 125: Perform hierarchical aggregation processing on the verified set of encrypted gradients, and reassemble the encrypted data according to the hierarchical order of the model hierarchical structure descriptor to generate global model encrypted data that has completed hierarchical aggregation encryption processing.

[0050] The system performs hierarchical aggregation processing on the verified set of encrypted gradients. Following the hierarchical order of the model's hierarchical structure descriptor, the encrypted data is reassembled. For each level, the encrypted gradient information units at that level are aggregated, for example, by combining encrypted gradient information units from multiple participants at the same level using a weighted average. Then, the aggregation results from each level are combined according to the model's hierarchical order to generate the global model encrypted data, which has undergone hierarchical aggregation encryption processing. During the reassembly process, it is ensured that the hierarchical relationship and structure of the data are consistent with the model's hierarchical structure descriptor, guaranteeing the accuracy and validity of the global model encrypted data.

[0051] Step 130: Distribute the global model encrypted data to each federated learning participant according to a preset fragmentation transmission strategy. The fragmentation transmission strategy is adjusted according to the network bandwidth and computing resource capacity of the federated learning participant.

[0052] After obtaining the globally encrypted model data that has undergone hierarchical aggregation and encryption processing, the data development and utilization system needs to distribute it to each federated learning participant. Due to differences in network bandwidth and computing resource capacity among the participants, a pre-defined fragmentation transmission strategy is adopted to ensure efficient and stable data transmission.

[0053] In a preferred embodiment, the step of distributing the global model encrypted data to each federated learning participant according to a preset fragmentation transmission strategy, wherein the fragmentation transmission strategy is adjusted based on the network bandwidth and computing resource capacity of the federated learning participant, includes:

[0054] Step 131: Collect real-time network bandwidth and computing resource capacity data of each federated learning participant, and generate a participant resource status vector.

[0055] The data development and utilization system collects real-time network bandwidth and computing resource capacity data through communication with each participant in the federated learning process. Network bandwidth reflects the participant's network transmission capability at the current moment, while computing resource capacity data includes information such as the participant's memory capacity and number of processor cores. The system integrates this data into a participant resource state vector, which contains information across multiple dimensions, including network bandwidth, memory capacity, and number of processor cores. For example, a participant resource state vector can be represented as (network bandwidth, memory capacity, number of processor cores).

[0056] Step 132: Construct a fragmented transmission decision model based on the resource state vector of the participants, using the total amount of encrypted data in the global model as the input variable, and the network bandwidth status and computing resource capacity as constraints.

[0057] A fragmented transmission decision model is constructed based on the collected resource state vectors of the participants. The purpose of this model is to determine the optimal number and size of fragments to meet the network bandwidth and computing resource constraints of each participant. The total amount of encrypted data in the global model is used as the input variable, and network bandwidth and computing resource capacity are used as constraints.

[0058] Optionally, the step of constructing a fragmented transmission decision model based on the resource state vectors of the participating parties, using the total amount of encrypted data in the global model as an input variable and network bandwidth status and computing resource capacity as constraints, includes:

[0059] Step 1321: Perform fluctuation interference optimization on the network bandwidth status in the resource state vector of the participating parties to obtain stable bandwidth feature values.

[0060] The network bandwidth status of participating parties may fluctuate. To obtain more accurate bandwidth information, fluctuation interference optimization is performed on the network bandwidth status in the participant's resource state vector. A moving average method is used to smooth the network bandwidth data, removing short-term fluctuations. By calculating the average network bandwidth over a period of time, a stable bandwidth characteristic value is obtained, which better reflects the actual network transmission capacity of the participating parties.

[0061] Step 1322: Normalize the memory capacity and processor core count in the computing resource capacity data to generate a comprehensive computing resource index.

[0062] Memory capacity and processor core count in computing resource capacity data have different dimensions. To consider them comprehensively, memory capacity and processor core count are normalized. A min-max normalization method is used to map memory capacity and processor core count to the interval [0, 1]. Then, the normalized memory capacity and processor core count are weighted and summed according to certain weights to generate a comprehensive computing resource index, which reflects the overall computing capability of the participants.

[0063] Step 1323: Using the total amount of encrypted data in the global model as the objective function, and the stable bandwidth characteristic value and the comprehensive index of computing resources as constraints, construct the mathematical programming expression of the segmented transmission decision model.

[0064] The objective function is to transmit the encrypted data of the global model as efficiently as possible while satisfying the network bandwidth and computing resource limitations of all participants. A mathematical programming expression for the fragmented transmission decision model is constructed, using stable bandwidth characteristic values ​​and a comprehensive computing resource index as constraints. This expression describes the relationship between the number of fragments, fragment size, and network bandwidth and computing resources.

[0065] Step 1324: Introduce a fragmentation transmission delay penalty factor into the mathematical programming expression. The delay penalty factor is positively correlated with the number of fragments and negatively correlated with the stable bandwidth characteristic value.

[0066] To avoid increased transmission delay due to excessive fragmentation, a fragmentation transmission delay penalty factor is introduced into the mathematical programming expression. This penalty factor is positively correlated with the number of fragments, meaning the more fragments, the larger the penalty factor; and negatively correlated with the stable bandwidth characteristic value, meaning the larger the stable bandwidth characteristic value, the smaller the penalty factor. By introducing this penalty factor, the impact of transmission delay is considered during the optimization process, making the model more reasonable.

[0067] Step 1325: Solve the mathematical programming expression using the Lagrange multiplier method to obtain a feasible solution set of the optimal number of partitions and partition size parameters that satisfy the constraints.

[0068] The Lagrange multiplier method is a technique for solving constrained optimization problems. It transforms the constraints into part of the objective function by introducing Lagrange multipliers. During the solution process, a set of feasible solutions with optimal number and size of partitions that satisfy the constraints is found. The optimal number and size of partitions are then selected from these feasible solutions to ensure efficient and stable data transmission.

[0069] Step 133: Calculate the optimal number of fragments and fragment size parameters for each federated learning participant using the fragmented transmission decision model.

[0070] Based on the constructed fragmented transmission decision model, the optimal number of fragments and fragment size parameters for each federated learning participant are calculated. Since each participant has different network bandwidth and computing resource capacity, the optimal number of fragments and fragment size parameters will also differ. For example, participants with higher network bandwidth and more abundant computing resources may receive larger fragment sizes and fewer fragments, while participants with lower network bandwidth and limited computing resources may receive smaller fragment sizes and more fragments.

[0071] Step 134: Perform fragmentation processing on the global model encrypted data according to the optimal fragmentation number and fragmentation size parameters to generate a set of encrypted data fragments containing fragmentation identifiers and sequential indexes.

[0072] Based on the calculated optimal number of fragments and fragment size parameters, the global model encrypted data is fragmented. The system divides the global model encrypted data into multiple fragments according to the fragment size. A unique fragment identifier is generated for each fragment, which contains information such as the model-level feature code, data block type identifier, and participant association marker. Simultaneously, a fragment order index system is constructed. By analyzing the feature dependencies of each fragment in the global model encrypted data, a doubly linked list structure index containing pointers to predecessor and successor fragments is generated. The fragment identifier, the doubly linked list structure index, and the actual fragment data are multi-dimensionally associated and encapsulated to form a set of encrypted data fragments containing a feature integrity verification field.

[0073] In an alternative embodiment, the step of fragmenting the global model encrypted data according to the optimal fragment number and fragment size parameters to generate a set of encrypted data fragments containing fragment identifiers and sequential indexes includes:

[0074] Step 1341: Perform feature boundary detection on the global model encrypted data according to the optimal number of shards, identify model structure feature blocks with semantic integrity in the global model encrypted data, and divide the model structure feature blocks into a preset number of logical shard units based on the natural boundaries of the model structure feature blocks.

[0075] Feature boundary detection is performed on the globally encrypted model data based on the optimal number of shards. By analyzing the feature distribution in the globally encrypted model data, semantically intact model structural feature blocks are identified. These feature blocks represent different functional parts of the model, such as different layers in a neural network model. A predetermined number of logical shard units are divided based on the natural boundaries of the model structural feature blocks, ensuring that each logical shard unit has a certain degree of semantic integrity.

[0076] Step 1342: Perform feature-aware cutting on each logical shard unit based on the shard size parameter, detect the feature density distribution through a sliding window, and perform cutting operations in areas where the feature density is lower than the density threshold to generate actual shard data.

[0077] Each logical partition unit is segmented using feature-aware cutting based on the partition size parameter. A sliding window method is employed to detect the feature density distribution. The sliding window moves across the logical partition unit, calculating the feature density within each window. When the feature density is below a density threshold, a segmentation operation is performed in that region, dividing the logical partition unit into multiple actual partition data. This ensures that each actual partition data contains sufficient feature information while meeting the partition size requirements.

[0078] Step 1343: Construct a unique shard identifier for each actual shard data. The shard identifier is generated by extracting the feature fingerprint of the shard data. The shard identifier includes the model-level feature code, the data block type identifier, and the participant association marker.

[0079] A unique shard identifier is constructed for each actual data shard. This identifier is generated by extracting the feature fingerprint of the shard data; the feature fingerprint is a unique representation of the shard data. The shard identifier includes information such as the model-level feature code, the data block type identifier, and the participant association marker. The model-level feature code indicates the model level to which the shard data belongs, the data block type identifier indicates the type of data block, and the participant association marker indicates which federated learning participant the shard data comes from.

[0080] Step 1344: Construct a sharding order index system. By analyzing the feature dependencies of each actual shard data in the global model encrypted data, generate a doubly linked list structure index containing predecessor shard pointers and successor shard pointers.

[0081] A fragment order indexing system is constructed to analyze the feature dependencies of each actual fragment in the globally encrypted model data. Based on these dependencies, a doubly linked list index containing pointers to predecessor and successor fragments is generated. The predecessor pointer points to the fragment preceding the current fragment, and the successor pointer points to the fragment following the current fragment. The doubly linked list index allows for easy determination of the fragment order and dependencies.

[0082] Step 1345: The shard identifier, doubly linked list structure index and actual shard data are associated and encapsulated in a multi-dimensional way to form an encrypted data shard set containing a feature integrity verification field. The feature integrity verification field is generated based on the feature fingerprint and sequential index relationship of the shard data.

[0083] The shard identifier, doubly linked list structure index, and actual shard data are encapsulated and associated in multiple dimensions. By integrating this information, a set of encrypted data shards containing a feature integrity verification field is formed. This feature integrity verification field is generated based on the feature fingerprint and sequential index relationship of the shard data and is used to verify the integrity and correctness of the shard data order. During transmission, checking the feature integrity verification field ensures that the shard data has not been tampered with or lost.

[0084] Step 135: Distribute the encrypted data fragment set to the corresponding federated learning participants by dynamically adjusting the transmission rate according to the sequential index of the encrypted data fragment set and the network bandwidth status of the federated learning participants.

[0085] The encrypted data fragments are distributed to the corresponding federated learning participants according to their sequential index. During transmission, the transmission rate is dynamically adjusted based on the network bandwidth of each participant. If a participant has high network bandwidth, the transmission rate is increased; if their network bandwidth is low, the transmission rate is decreased. This dynamic adjustment ensures efficient and stable data transmission. Simultaneously, transmission status, such as success rate and latency, is monitored in real time, and problems encountered during transmission, such as retransmitting lost fragments, are addressed promptly.

[0086] For example, the fragmented transmission decision model is an optimization model based on mathematical programming. Regarding input variables, the total amount of encrypted data in the global model is relatively large, for example, containing 10,000 data units. The network bandwidth of the participating parties is represented by the number of data units transmitted per second, ranging from 10 to 150 data units / second. In terms of computing resource capacity, memory capacity is represented by the number of data units that can be stored, ranging from 800 to 6000 data units, and the number of processor cores ranges from 1 to 10 cores.

[0087] When optimizing network bandwidth for fluctuations, the moving average method uses a moving window size of 7 time units to obtain stable bandwidth characteristic values. The memory capacity and processor core count in the computing resource capacity data are normalized to generate a comprehensive computing resource index, with memory capacity weighted at 0.65 and processor core count weighted at 0.35.

[0088] A mathematical programming expression is constructed with the total amount of encrypted data in the global model as the objective function, and the stable bandwidth characteristic value and the comprehensive index of computing resources as constraints. A fragmentation transmission delay penalty factor is introduced, with a correlation coefficient of 0.6 with the number of fragments and a correlation coefficient of -0.4 with the stable bandwidth characteristic value. The Lagrange multiplier method is used to solve the problem, with the initial value of the Lagrange multiplier set to 0.15 and 120 iterations.

[0089] Step 140: Guide each federated learning participant to use the global model encrypted data to update parameters and optimize the structure of the local business model, and generate optimization suggestion information by evaluating the model update status through the local validation dataset of each federated learning participant.

[0090] After the data development and utilization system distributes the globally encrypted model data to each federated learning participant, it guides each participant to use this data to update parameters and optimize the structure of their local business models. Each participant's local business model is trained based on their own non-shared raw data; by combining it with the globally encrypted model data, the model's performance can be further improved.

[0091] In one possible design, the guidance for each federated learning participant to use the global model encrypted data to update parameters and optimize the structure of their local business model, and to generate optimization suggestion information by evaluating the model update status using the local validation datasets of each federated learning participant, including:

[0092] Step 141: Send model co-optimization instructions to each federated learning participant. The model co-optimization instructions include the homomorphic decryption rules for global model encrypted data and the parameter feature association protocol.

[0093] The data development system sends model collaborative optimization instructions to all federated learning participants. These instructions include homomorphic decryption rules for the globally encrypted model data and a parameter feature association protocol. The homomorphic decryption rules guide participants in decrypting the globally encrypted model data, while the parameter feature association protocol specifies how to associate the decrypted global model parameter features with the parameter features of the local business model. For example, the parameter feature association protocol can specify feature matching based on the model's hierarchical structure or association based on the semantic information of the features.

[0094] Step 142: Receive the global parameter feature set extracted by each federated learning participant based on the decrypted global model encrypted data. The global parameter feature set includes model hierarchical features and gradient evolution features shared across participants.

[0095] After receiving the model collaborative optimization instruction, each participant in the federated learning process decrypts the encrypted global model data according to homomorphic decryption rules. Then, they extract the global parameter feature set from the decrypted global model data. The global parameter feature set contains model hierarchical features and gradient evolution features shared across participants. Model hierarchical features reflect the model's hierarchical structure and parameter distribution, while gradient evolution features reflect the changes in parameters during training.

[0096] Preferably, the step of receiving the global parameter feature set extracted by each federated learning participant based on the decrypted global model encrypted data includes:

[0097] Step 1421: Enable the federated learning participants to perform homomorphic decryption on the encrypted data of the global model through a preset decryption protocol, and separate the hierarchical parameter feature sequence and gradient evolution trajectory features of the global model.

[0098] Federated learning participants perform homomorphic decryption on the encrypted data of the global model using a pre-defined decryption protocol. Homomorphic decryption allows for the decryption of encrypted data without compromising data privacy. Through decryption, the hierarchical parameter feature sequences and gradient evolution trajectory features of the global model are separated. The hierarchical parameter feature sequences record the parameter values ​​at each level of the model, while the gradient evolution trajectory features record the changes in model parameters during training.

[0099] Step 1422: Enable the federated learning participants to discretize the hierarchical parameter feature sequence, mapping the continuous parameter space into a feature symbol sequence.

[0100] The hierarchical parameter feature sequence is discretized. Since the parameter values ​​in the hierarchical parameter feature sequence are continuous, to facilitate subsequent processing and analysis, the continuous parameter space is mapped to a sequence of feature symbols. A clustering algorithm is used to divide the continuous parameter values ​​into different categories, with each category represented by a feature symbol. For example, the K-means clustering algorithm is used to divide the parameter values ​​into K categories, with each category corresponding to a feature symbol.

[0101] Step 1423: Enable federated learning participants to extract time-series features of gradient evolution trajectory, identify the directional and magnitude correlations of gradient changes, and construct a gradient feature correlation map.

[0102] Time-series feature extraction is performed on gradient evolution trajectory features. By analyzing the time series of gradient evolution trajectory features, the directional and magnitude correlations of gradient changes are identified. Directional correlation reflects whether the direction of gradient changes is consistent, while magnitude correlation reflects the relationship between the magnitudes of gradient changes. Based on these correlations, a gradient feature correlation graph is constructed. The gradient feature correlation graph uses a graph structure to represent the relationships between gradients, with nodes representing gradients and edges representing the relationships between gradients.

[0103] Step 1424: Enable the federated learning participants to fuse the feature symbol sequence with the gradient feature association map to generate a global parameter feature set containing hierarchical structure features and evolutionary dynamic features.

[0104] Feature fusion is performed between the feature symbol sequence and the gradient feature association map. By integrating the information from both, a global parameter feature set containing hierarchical structure features and evolutionary dynamic features is generated. The hierarchical structure features are represented by the feature symbol sequence, and the evolutionary dynamic features are represented by the gradient feature association map. Feature fusion can be performed by concatenating the information from the feature symbol sequence and the gradient feature association map to form a new feature set.

[0105] Step 1425: Receive the global parameter feature set, provided that the global parameter feature set contains parameter feature components contributed by all federated learning participants.

[0106] The data development and utilization system receives the global parameter feature set, provided that it contains parameter feature components contributed by all participants in the federated learning process. Ensuring the integrity of the global parameter feature set is crucial for accurate subsequent analysis and processing. Missing parameter feature components from any participant may negatively impact model performance and optimization results.

[0107] Step 143: Guide each federated learning participant to perform feature association matching between the global parameter feature set and the current parameter feature set of the local business model, and identify complementary and conflicting regions in the feature space.

[0108] The data development system guides each participant in the federated learning process to perform feature association matching between the global parameter feature set and the current parameter feature set of their local business model. Through matching, complementary and conflicting regions in the feature space are identified. Complementary regions represent parts of the global and local parameter feature sets that can complement each other, while conflicting regions represent parts that contradict or overlap between the two.

[0109] Specifically, guiding each federated learning participant to perform feature association matching between the global parameter feature set and the current parameter feature set of the local business model, and identifying complementary and conflicting regions in the feature space, includes instructing the federated learning participants to perform the following operations:

[0110] Step 1431: Extract the current parameter feature set of the local business model, including the local hierarchical feature symbol sequence and the local gradient correlation graph.

[0111] Participants extract the current parameter feature set of the local business model, which includes a local hierarchical feature symbol sequence and a local gradient correlation graph. The local hierarchical feature symbol sequence is obtained by discretizing the hierarchical parameters of the local business model, and the local gradient correlation graph is obtained by feature extraction and correlation analysis of the gradient evolution trajectory of the local business model.

[0112] In this embodiment, the feature discretization process uses the K-means clustering algorithm, with K set to 6, dividing the continuous parameter space into 6 categories. The maximum number of iterations is set to 60, and the initial cluster centers are randomly selected.

[0113] Step 1432: Perform feature space alignment processing on the global parameter feature set and the local parameter feature set to unify the encoding rules of the feature symbol sequence and the topological structure of the gradient correlation graph.

[0114] Since the global parameter feature set and the local parameter feature set may employ different encoding rules and topological structures, it is necessary to unify the encoding rules of the feature symbol sequences and the topological structure of the gradient correlation graph in order to achieve accurate matching. For example, the feature symbol sequences of the global parameter feature set and the local parameter feature set can be encoded using the same method, and the topological structure of the gradient correlation graph can be adjusted to make it consistent.

[0115] Step 1433: The aligned feature set is compared layer by layer using the feature similarity calculation method to generate a feature similarity matrix, where the matrix elements represent the degree of correlation of features at the corresponding level.

[0116] The aligned feature sets are compared layer by layer using feature similarity calculation methods. Cosine similarity and other methods are used to calculate the similarity of the aligned feature sets at each layer. The calculated similarity values ​​are then used to form a feature similarity matrix, where each element represents the degree of association between features at the corresponding layer. A higher similarity value indicates a stronger association between features at the corresponding layer.

[0117] Step 1434: Based on the feature similarity matrix, identify feature regions with similarity higher than a preset threshold as complementary regions and feature regions with similarity lower than a preset threshold as conflict regions.

[0118] Complementary and conflicting regions are identified based on a feature similarity matrix. A preset threshold is set, classifying feature regions with similarity higher than the threshold as complementary regions and those with similarity lower than the threshold as conflicting regions. The preset threshold can be adjusted according to actual conditions to balance the division of complementary and conflicting regions.

[0119] Step 1435: Mark the feature boundaries of complementary and conflicting regions to generate feature region partitioning results that include region location identifiers and feature association types.

[0120] Complementary and conflicting regions are marked with feature boundaries to define their extent. The resulting feature region partitioning output includes region location identifiers and feature association types. Region location identifiers indicate the positions of complementary and conflicting regions in the feature space, while feature association types indicate whether a region is complementary or conflicting.

[0121] Step 144: Adjust the hierarchical correlation degree of the network structure of the local business model based on the feature association matching results. The hierarchical correlation degree adjustment includes strengthening the cross-level connection weight through feature complementary regions and suppressing redundant hierarchical nodes through conflict regions.

[0122] Based on feature association matching results, the hierarchical correlation of the network structure of the local business model is adjusted. For regions with complementary features, cross-level connection weights are strengthened to enhance the model's ability to utilize complementary information. For example, in a neural network model, the connection weights between layers corresponding to complementary regions are increased. For conflicting regions, redundant hierarchical nodes are suppressed to reduce the model's complexity and computational cost. For example, redundant nodes corresponding to conflicting regions are deleted or their connection weights are reduced.

[0123] Step 145: Obtain the feature response sequence obtained by each federated learning participant using the local validation dataset to test the feature response of the adjusted model. Based on the consistency and volatility analysis of the feature response sequence, generate optimization suggestion information including parameter feature adaptation suggestions and structural correlation optimization directions.

[0124] Each participant in the federated learning process uses a local validation dataset to test the feature response of the adjusted model. By inputting the local validation dataset, the output of each level of the model is recorded to obtain a feature response sequence. Based on the consistency and volatility analysis of the feature response sequence, optimization suggestions are generated, including parameter feature adaptation suggestions and structural correlation optimization directions.

[0125] For example, the step of obtaining the feature response sequence obtained by each federated learning participant using a local validation dataset to test the feature response of the adjusted model, and generating optimization suggestion information containing parameter feature adaptation suggestions and structural correlation optimization directions based on the consistency and volatility analysis of the feature response sequence, includes:

[0126] Step 1451: Instruct the federated learning participants to divide the local validation dataset into basic feature subsets, marginal feature subsets, and abnormal feature subsets according to feature type, and input them into the adjusted local business model in sequence. Collect the hierarchical feature response subsequences and cross-hierarchical correlation feature response subsequences output by each level of the model, and integrate them to form a feature response sequence that includes intra-hierarchical responses and inter-hierarchical correlations.

[0127] The federated learning participants are instructed to divide the local validation dataset into three subsets based on feature type: basic feature subset, marginal feature subset, and anomalous feature subset. The basic feature subset contains the main features of the dataset, the marginal feature subset contains some secondary, peripheral features, and the anomalous feature subset contains outlier data. These subsets are then sequentially input into the adjusted local business model, and hierarchical feature response subsequences and cross-hierarchical correlation feature response subsequences are collected from each level of the model. The hierarchical feature response subsequences reflect the model's output at each level, while the cross-hierarchical correlation feature response subsequences reflect the correlations between different levels. These subsequences are then integrated to form a feature response sequence that includes both intra-hierarchical responses and inter-hierarchical correlations.

[0128] Step 1452: Perform multi-scale temporal pattern analysis on the feature response sequence, identify the pattern evolution trend of the hierarchical feature response subsequences corresponding to continuous input samples in the first time window, detect the periodic correlation strength change of cross-hierarchical correlation feature response subsequences in the second time window, and generate a temporal pattern stability map.

[0129] Multi-scale temporal pattern analysis is performed on the feature response sequences. Within the first time window, the pattern evolution trend of hierarchical feature response subsequences corresponding to consecutive input samples is analyzed. By observing the changes in hierarchical feature response subsequences, the direction and trend of their pattern evolution are identified. Within the second time window, the periodic correlation strength changes of cross-hierarchical correlated feature response subsequences are detected. By analyzing the periodicity of cross-hierarchical correlated feature response subsequences, the changes in their correlation strength are determined. Based on these analysis results, a temporal pattern stability map is generated. The temporal pattern stability map graphically represents the temporal pattern stability of the feature response sequences, thereby identifying potential problems and optimization directions.

[0130] Step 1453: Perform spatial conduction path analysis on the feature response sequence, construct a directed feature conduction graph based on the activation order of the hierarchical feature response subsequences, identify feature blocking nodes and redundant conduction branches in the conduction path through graph structure analysis, and generate a spatial conduction efficiency evaluation matrix.

[0131] Spatial conduction path analysis is performed on the feature response sequences. A directed feature conduction graph is constructed based on the activation order of the hierarchical feature response subsequences. Nodes in the graph represent the hierarchy of the model, and edges represent the feature conduction relationships between levels. Through graph structure analysis, feature blocking nodes and redundant conduction branches in the conduction path are identified. Feature blocking nodes are nodes that hinder feature transmission during the conduction process, while redundant conduction branches are branches that do not contribute substantially to feature conduction. Based on these analysis results, a spatial conduction efficiency evaluation matrix is ​​generated. The spatial conduction efficiency evaluation matrix represents the efficiency of feature conduction in matrix form, which can be used to identify problems in the spatial conduction path.

[0132] Step 1454: Construct a multi-dimensional feature response quality assessment model by combining the temporal pattern stability map and the spatial transmission efficiency evaluation matrix. The multi-dimensional feature response quality assessment model includes a temporal consistency dimension, a spatial transmission dimension, and a feature robustness dimension. The feature robustness dimension is determined by analyzing the feature response decay rate and recovery capability corresponding to the abnormal feature subset.

[0133] A multi-dimensional feature response quality assessment model is constructed by combining temporal pattern stability maps and spatial transmission efficiency evaluation matrices. This model includes temporal consistency, spatial transmission efficiency, and feature robustness dimensions. The temporal consistency dimension reflects the consistency of the feature response sequence over time; the spatial transmission efficiency reflects the transmission efficiency of features within the model; and the feature robustness dimension is determined through the analysis of feature response decay rates and recovery capabilities corresponding to anomalous feature subsets. The feature response decay rate corresponding to anomalous feature subsets reflects the degree of performance degradation of the model when encountering anomalous data, while the recovery capability reflects the model's ability to recover normal performance after the influence of anomalous data disappears.

[0134] Step 1455: Output the feature response quality index of each dimension through the multi-dimensional feature response quality assessment model, and identify the temporal fragile region with a temporal consistency index lower than the benchmark value, the conduction stagnation region with a spatial conduction efficiency index lower than the benchmark value, and the anti-interference weak region with a robustness index lower than the benchmark value.

[0135] The multi-dimensional feature response quality assessment model outputs feature response quality indices for each dimension. The temporal consistency index reflects the consistency of the feature response sequence over time, the spatial transmission efficiency index reflects the transmission efficiency of features within the model, and the robustness index reflects the model's resistance to interference. It identifies temporally vulnerable regions with temporal consistency indices below the benchmark, regions with transmission stagnation regions with spatial transmission efficiency indices below the benchmark, and regions with weak resistance to interference with robustness indices below the benchmark. The benchmark values ​​can be set according to actual conditions to determine the model's performance standards across each dimension.

[0136] Step 1456: For temporally vulnerable regions, generate parameter feature temporal alignment suggestions based on pattern evolution trend analysis, including feature response delay compensation mechanisms and cross-window pattern association weight adjustment strategies; for regions with conduction stagnation, generate structural association optimization directions based on path analysis of the directed graph of feature conduction, including strengthening the adjacency-level connections of stagnation nodes and node reduction schemes for redundant branches; for regions with weak anti-interference capabilities, generate feature enhancement and adaptation suggestions based on the response decay characteristics of abnormal feature subsets, including activation function type replacement and feature mapping space expansion strategies for abnormal feature sensitive levels.

[0137] For temporally vulnerable regions, parameter feature temporal alignment suggestions are generated based on pattern evolution trend analysis. These suggestions include a feature response delay compensation mechanism and a cross-window pattern association weight adjustment strategy. The feature response delay compensation mechanism compensates for the temporal delay of feature responses, while the cross-window pattern association weight adjustment strategy adjusts the pattern association weights between different time windows to improve temporal consistency. For regions with transmission stagnation, structural association optimization directions are generated based on path analysis of the directed feature transmission graph. These directions include strengthening the adjacency-level connections of stagnant nodes and reducing redundant branches. By strengthening the adjacency-level connections of stagnant nodes and reducing redundant branches, spatial transmission efficiency is improved. For regions with weak anti-interference capabilities, feature enhancement and adaptation suggestions are generated by combining the response decay characteristics of anomalous feature subsets. These suggestions include replacing the activation function type of anomalous feature-sensitive layers and expanding the feature mapping space. By replacing the activation function type of anomalous feature-sensitive layers and expanding the feature mapping space, the model's anti-interference capability is improved.

[0138] In the construction of the directed graph for feature propagation, graph nodes represent model levels; if the model has 6 levels, there are 6 nodes. Edges represent feature propagation relationships between levels, and edge weights are determined by the activation strength of the feature response subsequences of each level; the greater the activation strength, the greater the weight. Feature blocking nodes and redundant propagation branches are identified through graph structure analysis.

[0139] Step 1457: Based on temporal alignment suggestions, structural correlation optimization directions, and feature enhancement adaptation suggestions, generate optimization suggestion information that includes multi-dimensional optimization priority ranking.

[0140] Based on temporal alignment suggestions, structural correlation optimization directions, and feature enhancement adaptation suggestions, optimization suggestion information with multi-dimensional optimization priority ranking is generated. These suggestions are prioritized according to their importance and urgency in improving model performance. The multi-dimensional optimization priority ranking considers factors such as temporal consistency, spatial transport, and feature robustness to ensure effective optimization of the model in multiple aspects.

[0141] In this embodiment, the multi-dimensional feature response quality assessment model uses a multilayer perceptron (MLP) to evaluate the feature response quality. The number of neurons in the input layer depends on the dimension of the feature response sequence. If it includes 12 dimensions of features, including intra-layer response and inter-layer correlation, the number of neurons in the input layer is 12, and the input data is a 12-dimensional vector.

[0142] The first hidden layer has 30 neurons, activated by ReLU. The second hidden layer has 20 neurons, also activated by ReLU. The output layer has 3 neurons, corresponding to the feature response quality indices for temporal consistency, spatial transmissibility, and feature robustness, respectively. The activation function is a linear function directly outputting the indices. Training is performed using the Adagrad optimization algorithm with a learning rate of 0.008. The loss function is the mean squared error loss, and the training run consists of 220 epochs.

[0143] In an extended embodiment, the method further includes:

[0144] Step 210: Before each federated learning participant uploads the encrypted model data, guide the federated learning participants to perform feature desensitization preprocessing on the local non-shared raw data. Remove identity features and sensitive attribute features from the raw data through feature abstraction processing, and retain the structural features and correlation features required for model training.

[0145] Before each participant in the federated learning process uploads encrypted model data, they are guided to perform feature desensitization preprocessing on their local, non-shared raw data. The purpose of feature desensitization preprocessing is to protect data privacy and security while retaining the effective information needed for model training. This involves removing identity-identifying and sensitive attribute features from the raw data through feature abstraction. Identity-identifying features include a user's name and ID number, while sensitive attribute features include a user's health and financial information. Structural and relational features needed for model training are retained, such as the relationship between raw material procurement and production process parameters, and the relationship between sales channels and sales volume. Feature abstraction can employ methods such as generalization and masking. Generalization methods replace specific feature values ​​with more abstract categories, such as replacing specific age values ​​with age ranges; masking methods replace sensitive feature values ​​with mask symbols, such as replacing parts of an ID number with asterisks.

[0146] Step 220: After receiving the de-identified encrypted model data, perform reversibility verification on the feature abstraction results in the data to ensure that the de-identification process does not destroy the key feature correlation required for model training. After the verification is passed, perform hierarchical aggregation encryption processing.

[0147] After receiving the de-identified encrypted model data, the reversibility of the feature abstraction results in the data is verified. The purpose of reversibility verification is to ensure that the de-identification process has not destroyed the key feature correlations required for model training. By performing a reverse operation on the feature abstraction results, it is checked whether the key feature correlations of the original data can be recovered. For example, for features after generalization, it is checked whether the approximate range of the original features can be restored according to the generalization rules; for features after masking, it is checked whether the key feature correlations can be preserved without disclosing sensitive information. If the verification passes, hierarchical aggregation encryption processing is performed, as described in step 120.

[0148] Step 230: During the global model encrypted data distribution phase, a unique feature mask is generated for each federated learning participant. The mask is dynamically generated based on the participant's de-identified feature distribution and is used to mask the de-identified features of other federated learning participants that are unrelated to that participant.

[0149] During the global model encrypted data distribution phase, a unique feature mask is generated for each federated learning participant. This unique feature mask is dynamically generated based on the participant's anonymized feature distribution. By analyzing the participant's anonymized feature distribution, it is determined which features are relevant to that participant and which are irrelevant. The unique feature mask is used to mask the anonymized features of other federated learning participants unrelated to that participant, ensuring that the participant can only obtain feature information relevant to its own business. For example, for a manufacturer, the unique feature mask can mask sensitive feature information of distributors and retailers, retaining only feature information relevant to production.

[0150] Step 240: After receiving the encrypted global model data containing the exclusive feature mask, the federated learning participants filter out irrelevant features through the mask and retain the feature components related to the local business model for parameter updates and structural optimization.

[0151] After receiving encrypted global model data containing a proprietary feature mask, the federated learning participants filter out irrelevant features using the mask. Each participant uses the proprietary feature mask to filter the encrypted global model data, retaining only feature components relevant to their local business model. Then, these retained feature components are used to update the parameters and optimize the structure of the local business model, as described in step 140. Filtering irrelevant features through the mask ensures that participants only use information relevant to their own business when updating and optimizing the model, improving the model's relevance and accuracy.

[0152] Step 250: Verify whether sensitive features are effectively isolated through simulated attack tests, jointly evaluate the feature desensitization effect and mask filtering effect, and adjust the feature desensitization rules and mask generation strategy according to the evaluation results to maintain a dynamic balance between privacy protection and model performance.

[0153] The effectiveness of sensitive features isolation was verified through simulated attack tests. These tests employed various attack methods, such as data mining and machine learning attacks, to attempt to recover sensitive feature information from the de-identified encrypted model data and filtered feature components. If the sensitive feature information could not be recovered in the simulated attack tests, it indicated that the sensitive features were effectively isolated. The effectiveness of feature de-identification and masking filtering was jointly evaluated. Evaluation metrics included the degree of privacy protection and the degree of model performance loss. The degree of privacy protection reflected the extent to which sensitive features were protected, while the degree of model performance loss reflected the impact of feature de-identification and masking filtering on model performance. Based on the evaluation results, the feature de-identification rules and masking generation strategies were adjusted. If the degree of privacy protection was insufficient, the feature de-identification rules and masking generation strategies were strengthened; if the model performance loss was too large, the feature de-identification rules and masking generation strategies were appropriately relaxed to maintain a dynamic balance between privacy protection and model performance.

[0154] In an extended embodiment, the method further includes:

[0155] Step 310: Construct a federated learning feature interaction behavior feature sequence library, which records the feature interaction behavior feature sequences of each federated learning participant in the process of uploading, downloading and updating encrypted model data and parameters. The behavior feature sequence includes interaction type identifier, model level association features, feature flow vector and interaction context association features.

[0156] A feature sequence library of interactive behaviors in federated learning is constructed. This library records the feature sequences of interactive behaviors of each participant in the process of uploading, downloading, and updating encrypted model data. The behavioral feature sequences include interaction type identifiers, model hierarchy-related features, feature flow vectors, and interaction context-related features. The interaction type identifier indicates the type of interaction, such as upload, download, or parameter update; the model hierarchy-related features indicate the model hierarchy involved in the interaction; the feature flow vector indicates the direction of feature flow; and the interaction context-related features indicate the context information of the interaction, such as time and location. By recording these feature sequences of interactive behaviors, the behavior of the participants can be analyzed and monitored, and abnormal behaviors can be detected in a timely manner.

[0157] Step 320: Extract contextual features from the feature sequence of interactive behaviors using a temporal correlation algorithm, establish a cross-participant feature interaction behavior correlation graph, and identify feature interaction paths with abnormal connection strength in the graph.

[0158] Contextual features are extracted from feature interaction behavior sequences using a temporal correlation algorithm. This algorithm analyzes the temporal order and relationships of feature interaction behavior sequences to extract contextual features. Based on these contextual features, a cross-partner feature interaction behavior correlation graph is constructed. Nodes in the graph represent the feature interaction behaviors of the participants, and edges represent the relationships between these behaviors. By analyzing the connection strength of the edges in the graph, feature interaction paths with abnormal connection strength are identified. Abnormal connection strength may indicate abnormal feature interaction behaviors, such as data leakage or malicious attacks.

[0159] Step 330: Input the feature sequence of interactive behavior into the pre-trained temporal anomaly detection model, and use the model's feature attention mechanism module to weight the interactive context-related features to generate the probability distribution of abnormal behavior of the feature interactive behavior.

[0160] The feature sequences of interactive behaviors are input into a pre-trained temporal anomaly detection model. This model employs structures such as Recurrent Neural Networks (RNNs) or Long Short-Term Memory Networks (LSTMs) in deep learning, possessing temporal processing capabilities. The model's feature attention mechanism module weights the interaction context-related features, highlighting important context-related features. Through the model's processing, an anomalous behavior probability distribution of the feature interactions is generated, representing the probability that each feature interaction is an anomalous behavior.

[0161] Step 340: Based on the probability distribution of abnormal behavior, filter out the feature interaction behavior segments whose abnormal probability value exceeds the set probability value, and perform feature mutation pattern analysis on the model hierarchical association features and feature flow vectors in the feature interaction behavior segments to obtain the abnormal behavior segments.

[0162] Based on the probability distribution of abnormal behavior, feature interaction behavior segments with abnormal probability values ​​exceeding a set probability value are selected. A probability threshold is set, and feature interaction behavior segments with abnormal probability values ​​exceeding this threshold are then selected. Feature mutation pattern analysis is performed on the model hierarchical association features and feature flow vectors in these abnormal behavior segments. Feature mutation pattern analysis can discover the change patterns of features in abnormal behavior segments, such as abrupt changes in feature values ​​and anomalous feature flows. Based on the analysis results, abnormal behavior segments are obtained, and the specific characteristics and manifestations of abnormal behavior are clarified.

[0163] Step 350: Combining feature gene fingerprint extraction, interaction path topology analysis, and participant behavior baseline comparison, determine the risk level and potential threat type of the abnormal behavior fragment to obtain risk assessment results; generate dynamic response strategies based on risk level and potential threat type, including feature isolation strategy, interaction rate limiting strategy, and path redirection strategy.

[0164] By combining feature fingerprint extraction, interaction path topology analysis, and participant behavior baseline comparison, the risk level and potential threat type of anomalous behavior segments are determined. Feature fingerprint extraction identifies unique characteristics of anomalous behavior segments, interaction path topology analysis analyzes the propagation path of anomalous behavior, and participant behavior baseline comparison compares the differences between anomalous behavior and normal participant behavior. Based on these analysis results, the risk level and potential threat type of anomalous behavior segments are determined, such as low risk, medium risk, high risk, data leakage, malicious attack, etc. Dynamic response strategies are generated based on the risk level and potential threat type. Dynamic response strategies include feature isolation strategies, interaction rate limiting strategies, and path redirection strategies. Feature isolation strategies isolate the features involved in anomalous behavior to prevent its propagation; interaction rate limiting strategies restrict the interaction traffic of anomalous behavior to reduce its impact; and path redirection strategies change the interaction path of anomalous behavior, guiding it away from sensitive data and critical systems.

[0165] For example, the feature gene fingerprint extraction uses the MD5 hash algorithm to process abnormal behavior fragments and generate a 128-bit hash value as the feature gene fingerprint. When determining the risk level and potential threat type of abnormal behavior, similar abnormal behaviors are identified by comparing fingerprints.

[0166] Step 360: The abnormal behavior fragment, the risk assessment result, and the execution effect of the dynamic response strategy are used as security knowledge metadata and fed back to the feature interaction behavior feature sequence library to update the detection parameters of the time-series anomaly detection model and the feature weight coefficients of the multi-dimensional feature tracing analysis.

[0167] Abnormal behavior fragments, risk assessment results, and the effectiveness of dynamic response strategy execution are used as security knowledge metadata and fed back to the feature sequence library of interactive behaviors. This security knowledge metadata can be used to update the detection parameters of the time-series anomaly detection model and the feature weight coefficients of multi-dimensional feature tracing analysis. By updating the detection parameters, the detection accuracy of the time-series anomaly detection model is improved; by updating the feature weight coefficients, the effectiveness of multi-dimensional feature tracing analysis is optimized, enabling better detection and handling of abnormal behaviors. Simultaneously, the continuous accumulation of security knowledge metadata enhances the system's security and reliability.

[0168] In this embodiment, the temporal anomaly detection model uses a Long Short-Term Memory (LSTM) network to monitor feature interaction behavior. The number of neurons in the input layer is determined based on the feature sequence dimension of the feature interaction behavior. If it includes four dimensions of information—interaction type identifier, model hierarchical association features, feature flow vector, and interaction context association features—then the number of neurons in the input layer is 4, and the input data is a 4-dimensional vector.

[0169] The LSTM layer has 48 neurons, and long-term dependencies in the sequence data are handled through forget gates, input gates, and output gates. The fully connected layer has 1 neuron, and the sigmoid activation function outputs the probability of abnormal behavior. Training uses the Adam optimization algorithm with a learning rate of 0.0002. The binary cross-entropy loss function is used, and the training run is 250 epochs.

[0170] Please see Figure 2 The figure is a schematic diagram of the basic structure of a data development and utilization system 200 provided in an embodiment of this application. The data development and utilization system 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the data development and utilization methods based on the federated learning framework.

[0171] Please see Figure 3 This application provides a functional block diagram of a data development and utilization device based on a federated learning framework. The data development and utilization device based on a federated learning framework includes:

[0172] The data receiving module is used to receive encrypted model data transmitted by each federated learning participant through a secure communication link. The encrypted model data is generated by each federated learning participant after completing model training locally based on non-shared raw data.

[0173] The data encryption module is used to perform homomorphic encryption operation on the model parameter gradient information in the encrypted model data using an asymmetric encryption algorithm, and to obtain global model encrypted data after verification by combining the model structure metadata in the encrypted model data and completing the hierarchical aggregation encryption process.

[0174] The data distribution module is used to distribute the encrypted global model data to each federated learning participant according to a preset fragmentation transmission strategy. The fragmentation transmission strategy is adjusted according to the network bandwidth and computing resource capacity of the federated learning participant.

[0175] The federated training module guides each federated learning participant to update parameters and optimize the structure of their local business model using the global model encrypted data, and generates optimization suggestions by evaluating the model update status using the local validation datasets of each federated learning participant.

[0176] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0177] In summary, please refer to the following: Figure 4From a data utilization perspective, this application's embodiments break down the barriers of data fragmentation and reluctance to share in production, sales, and marketing by receiving encrypted model data generated by various federated learning participants based on non-shared raw data. This allows all parties to leverage multi-party data to extract more value while protecting data security and privacy. For example, combining data from the production and sales stages can provide enterprises with more accurate production and sales decision-making support.

[0178] In terms of model processing, this application embodiment performs homomorphic encryption operation on the model parameter gradient information in the encrypted model data and combines it with model structure metadata verification to complete the hierarchical aggregation encryption process to obtain global model encrypted data. This not only ensures the security of the data during the aggregation process, but also effectively integrates the advantages of each model to improve the performance and accuracy of the global model.

[0179] In terms of data transmission, this application embodiment adjusts the fragmentation transmission strategy according to the network bandwidth and computing resource capacity of the federated learning participants, ensuring that the global model encrypted data can be efficiently and stably distributed to each participant, avoiding data transmission bottlenecks caused by differences in network and computing resources, and improving the overall efficiency and adaptability of the system.

[0180] In the model optimization stage, the embodiments of this application guide each participant to use global model encrypted data to update parameters and optimize the structure of local business models, and generate optimization suggestion information through evaluation of local verification datasets, so that the local business models of each participant can be continuously optimized, realizing the whole process optimization from data to model to business.

[0181] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the data development and utilization system reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the data development and utilization system to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0182] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A data exploitation method based on a federated learning framework, characterized in that, The method is implemented by a data development and utilization system, the data development and utilization system is in communication connection with a plurality of federated learning participants, and the method comprises: Receiving encrypted model data transmitted by each federated learning participant through a secure communication link, the encrypted model data being generated by each federated learning participant after completing model training locally based on non-shared original data; Performing homomorphic encryption operation on model parameter gradient information in the encrypted model data through an asymmetric encryption algorithm, and obtaining global model encrypted data after completing layered aggregation encryption processing through verification in combination with model structure metadata in the encrypted model data; the model structure metadata comprises a hierarchical structure and parameter dimensions of the model; Distributing the global model encrypted data to each federated learning participant according to a preset sharding transmission strategy, the sharding transmission strategy being adjusted according to network bandwidth conditions and computing resource capacity of the federated learning participants; Guiding each federated learning participant to perform parameter updating and structure optimization on a local business model by using the global model encrypted data, and generating optimization suggestion information through model updating state evaluation by using a local validation data set of each federated learning participant: sending a model collaborative optimization instruction to each federated learning participant, the model collaborative optimization instruction comprising a homomorphic decryption rule of the global model encrypted data and a parameter feature association protocol; the parameter feature association protocol specifies how to associate the decrypted global model parameter features with parameter features of the local business model; receiving a global parameter feature set extracted by each federated learning participant based on the decrypted global model encrypted data, the global parameter feature set comprising model hierarchical features and gradient evolution features shared across the participants; the model hierarchical features reflect the hierarchical structure and parameter distribution of the model, and the gradient evolution features reflect changes in the parameters of the model during the training process; guiding each federated learning participant to perform feature association matching on the global parameter feature set and a current parameter feature set of the local business model, identify complementary regions and conflict regions in the feature space; the complementary regions represent parts that complement each other between the global parameter feature set and the local parameter feature set, and the conflict regions represent parts that have contradictions or repetitions between the global parameter feature set and the local parameter feature set; performing hierarchical association degree adjustment on the network structure of the local business model based on the feature association matching result, the hierarchical association degree adjustment comprising strengthening cross-hierarchical connection weights through the feature complementary regions and suppressing redundant hierarchical nodes through the conflict regions; obtaining a feature response sequence obtained by each federated learning participant through feature response testing on the adjusted model using a local validation data set, and generating optimization suggestion information comprising parameter feature adaptation suggestions and structure association optimization directions based on consistency and volatility analysis of the feature response sequence.

2. The method of claim 1, wherein, The method comprises: Extracting model structure metadata in the encrypted model data, and parsing to obtain a model hierarchical structure descriptor and a parameter dimension distribution feature; Performing hierarchical division processing on the model parameter gradient information based on the model hierarchical structure descriptor, to generate gradient information subsets corresponding to model levels; Performing homomorphic encryption operation on each gradient information subset by using a public key of an asymmetric encryption algorithm, to generate encrypted gradient information units; Associating the encrypted gradient information units with the corresponding parameter dimension distribution features, verifying the matching of the parameter dimensions before and after encryption, and obtaining an encrypted gradient set that passes the verification; Performing hierarchical aggregation processing on the encrypted gradient set that passes the verification, and recombining the encrypted data according to the hierarchical order of the model hierarchical structure descriptor, to generate global model encrypted data that has completed hierarchical aggregation encryption processing.

3. The method of claim 1, wherein, The global model encrypted data is distributed to each federated learning participant according to a preset sharding transmission strategy, and the sharding transmission strategy is adjusted according to the network bandwidth status and the computing resource capacity of the federated learning participant, and includes: Collecting real-time network bandwidth status and computing resource capacity data of each federated learning participant, and generating a participant resource state vector; Building a sharding transmission decision model based on the participant resource state vector, taking the total data volume of the global model encrypted data as an input variable, and taking the network bandwidth status and the computing resource capacity as constraint conditions; Calculating the optimal sharding number and sharding size parameters for each federated learning participant through the sharding transmission decision model; Sharding the global model encrypted data according to the optimal sharding number and sharding size parameters, to generate an encrypted data shard set containing shard identifiers and sequence indexes; According to the sequence indexes of the encrypted data shard set, dynamically adjusting the transmission rate in combination with the network bandwidth status of the federated learning participant, and distributing the encrypted data shard set to the corresponding federated learning participant.

4. The method of claim 3, wherein, Building a sharding transmission decision model based on the participant resource state vector, taking the total data volume of the global model encrypted data as an input variable, and taking the network bandwidth status and the computing resource capacity as constraint conditions, includes: Optimizing the network bandwidth status in the participant resource state vector for fluctuation interference, to obtain a stable bandwidth feature value; Normalizing the memory capacity and the number of processor cores in the computing resource capacity data, to generate a computing resource comprehensive index; Taking the total data volume of the global model encrypted data as a target function, and taking the stable bandwidth feature value and the computing resource comprehensive index as constraint conditions, constructing a mathematical programming expression of the sharding transmission decision model; Introducing a sharding transmission delay penalty factor in the mathematical programming expression, and the delay penalty factor has a positive correlation with the sharding number and a negative correlation with the stable bandwidth feature value; Solving the mathematical programming expression by the Lagrange multiplier method, to obtain a feasible solution set of the optimal sharding number and sharding size parameters that satisfy the constraint conditions.

5. The method of claim 4, wherein, The sharding processing of the global model encrypted data according to the optimal sharding number and sharding size parameters, to generate an encrypted data shard set containing shard identifiers and sequence indexes, includes: According to the optimal number of fragments, feature boundary detection is performed on the global model encrypted data, model structure feature blocks with semantic integrity in the global model encrypted data are identified, and a preset number of logical fragment units are divided based on the natural boundaries of the model structure feature blocks; Based on the fragment size parameter, each logical fragment unit is feature-aware cut, feature density distribution is detected through a sliding window, and cutting operations are performed in areas where the feature density is lower than the density threshold to generate actual fragment data; A unique fragment identifier is constructed for each actual fragment data, the fragment identifier is generated by extracting the feature fingerprint of the fragment data, and the fragment identifier includes a model level feature code, a data block type identifier, and a participant association marker; A fragment order index system is constructed, a bidirectional linked list structure index containing predecessor fragment pointers and successor fragment pointers is generated by analyzing the feature dependency relationship of each actual fragment data in the global model encrypted data; The fragment identifier, bidirectional linked list structure index, and actual fragment data are multi-dimensionally associated and packaged to form an encrypted data fragment set containing a feature integrity check field, which is generated based on the feature fingerprint and order index relationship of the fragment data.

6. The method of claim 1, wherein, The receiving of the global parameter feature set extracted by each federated learning participant based on the decrypted global model encrypted data includes: Causing the federated learning participant to perform a homomorphic decryption operation on the global model encrypted data through a preset decryption protocol to separate the hierarchical parameter feature sequence and the gradient evolution trajectory feature of the global model; Causing the federated learning participant to perform feature discretization processing on the hierarchical parameter feature sequence to map the continuous parameter space to a feature symbol sequence; Causing the federated learning participant to perform time series feature extraction on the gradient evolution trajectory feature to identify the direction correlation and amplitude correlation of the gradient change and construct a gradient feature correlation graph; Causing the federated learning participant to perform feature fusion on the feature symbol sequence and the gradient feature correlation graph to generate a global parameter feature set containing hierarchical structure features and evolution dynamic features; Receiving the global parameter feature set on the premise that the global parameter feature set contains parameter feature components contributed by all federated learning participants; The guiding of the federated learning participant to perform feature correlation matching on the global parameter feature set and the current parameter feature set of the local business model to identify complementary regions and conflict regions in the feature space includes: Causing the federated learning participant to perform the following operations: Extracting the current parameter feature set of the local business model, including a local hierarchical feature symbol sequence and a local gradient correlation graph; Performing feature space alignment processing on the global parameter feature set and the local parameter feature set to unify the encoding rules of the feature symbol sequence and the topology of the gradient correlation graph; Performing layer-by-layer comparison on the aligned feature sets through a feature similarity calculation method to generate a feature similarity matrix, and the matrix elements represent the correlation degree of the corresponding hierarchical features; Identifying feature regions with a similarity higher than a preset threshold as complementary regions and feature regions with a similarity lower than a preset threshold as conflict regions based on the feature similarity matrix. The complementary region and the conflict region are marked with feature boundary markers to generate a feature region division result containing region position identification and feature association type.

7. The method of claim 1, wherein, The feature response sequence obtained by performing feature response testing on the adjusted model using the local validation dataset of each federated learning participant is analyzed based on consistency and volatility to generate optimization suggestion information containing parameter feature adaptation suggestions and structure association optimization directions, including: The local validation dataset is divided into a basic feature subset, an edge feature subset, and an abnormal feature subset according to feature types, and the adjusted local business model is input in turn to collect the hierarchical feature response subsequence and the cross-hierarchical association feature response subsequence output by each level of the model, and integrate to form a feature response sequence containing intra-level response and inter-level association; Multi-scale time sequence pattern analysis is performed on the feature response sequence to identify the pattern evolution trend of the hierarchical feature response subsequence corresponding to the continuous input sample in the first time window, and detect the periodic association strength change of the cross-hierarchical association feature response subsequence in the second time window to generate a time sequence pattern stability graph; Spatial conduction path analysis is performed on the feature response sequence, a feature conduction directed graph is constructed based on the activation order of the hierarchical feature response subsequence, and feature blocking nodes and redundant conduction branches in the conduction path are identified through graph structure analysis to generate a spatial conduction efficiency evaluation matrix; A multi-dimensional feature response quality evaluation model is constructed by combining the time sequence pattern stability graph and the spatial conduction efficiency evaluation matrix, and the multi-dimensional feature response quality evaluation model contains a time sequence consistency dimension, a spatial conduction dimension, and a feature robustness dimension, and the feature robustness dimension is determined by analyzing the feature response decay rate and recovery ability of the abnormal feature subset; The multi-dimensional feature response quality evaluation model outputs feature response quality indexes of each dimension to identify time sequence fragile regions with a time sequence consistency index below a baseline value, conduction blocking regions with a spatial conduction efficiency index below a baseline value, and anti-interference weak regions with a robustness index below a baseline value; For the time sequence fragile region, a parameter feature time sequence alignment suggestion including a feature response delay compensation mechanism and a cross-window pattern association weight adjustment strategy is generated based on pattern evolution trend analysis; for the conduction blocking region, a structure association optimization direction including adjacent level connection strengthening of blocking nodes and node deletion scheme of redundant branches is generated based on path analysis of the feature conduction directed graph; for the anti-interference weak region, a feature enhancement adaptation suggestion including abnormal feature sensitive level activation function type replacement and feature mapping space expansion strategy is generated based on the response decay characteristics of the abnormal feature subset; Based on the time sequence alignment suggestion, the structure association optimization direction, and the feature enhancement adaptation suggestion, optimization suggestion information containing multi-dimensional optimization priority sorting is generated.

8. A data exploitation system, characterized by It includes: A processor, a storage device having a computer program stored thereon, and a network interface for providing network communication functions; when the computer program is executed by the processor, the processor implements the data development and utilization method based on the federated learning framework as claimed in any one of claims 1-7.

9. A readable storage medium, characterized by, The readable storage medium stores programs or instructions, and the programs or instructions are executed by the processor to implement the data development and utilization method based on the federated learning framework in any one of claims 1-7.

Images