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

Abstract

The embodiment of the invention discloses a data development and utilization method and system based on a federated learning framework. The method comprises the following steps: receiving encrypted model data which is transmitted by each federated learning participant through a secure communication link and is generated based on local non-shared original data training; performing homomorphic encryption operation on model parameter gradient information in the encryption model data, and combining model structure metadata verification to obtain global model encryption data after hierarchical aggregation encryption processing; distributing the global model encrypted data to each participant according to a fragmentation transmission strategy adjusted according to the network bandwidth condition of the participant and the computing resource capacity; according to the method, all participants are guided to perform parameter updating and structure optimization on a local service model by using global model encryption data, model updating state evaluation is performed through a local verification data set to generate optimization suggestion information, and the method can realize multi-party data cooperation sharing and improve model performance on the premise of protecting data security and privacy.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of computer data processing, and particularly relates to a data development and utilization method and system based on a federated learning framework. BACKGROUND

[0002] In the spicy seasoning application scenario, a large amount of data is generated in multiple links such as production, sales, and marketing. The production link covers detailed information of raw material procurement, such as the source, quality, and price of different raw materials, as well as production process parameters such as temperature, pressure, and time control; the sales link includes the diversity of sales channels, such as online e-commerce platforms, offline supermarkets, and convenience stores, as well as sales data of each channel; the marketing link has consumer feedback, including evaluation of product taste, packaging, and price, as well as market research information such as consumer preferences and competitor product conditions. These data are respectively mastered by different subjects such as producers, distributors, and retailers.

[0003] Traditional data sharing and utilization methods often require direct exchange of raw data, but due to data security and privacy protection considerations, parties are reluctant to directly share data. In the aspect of machine learning model training, the centralized training method is commonly used, that is, all data are centralized to a central node for training. However, this method faces many challenges in the spicy seasoning scenario because the data characteristics and distribution of different participants are quite different, and centralized training cannot fully utilize the advantages of data from each party, resulting in poor model performance. Moreover, direct sharing of raw data poses serious data security and privacy risks, which may leak enterprise business secrets.

[0004] In summary, the traditional centralized training method cannot adapt to the characteristics of scattered and different data in the spicy seasoning scenario, resulting in limited model performance; in addition, direct sharing of raw data poses serious data security and privacy risks, making it difficult for each party to achieve effective data cooperation and model training. SUMMARY

[0005] The embodiment of the present application provides a data development and utilization method and system based on a federated learning framework.

[0006] The embodiment of the present application provides a data development and utilization method based on a federated learning framework, applied to a data development and utilization system, and the method comprises the following steps: 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 based on non-shared raw data locally; The model parameter gradient information in the encrypted model data is homomorphically encrypted by an asymmetric encryption algorithm, and global model encrypted data subjected to hierarchical aggregation encryption processing is obtained after verification in combination with model structure metadata in the encrypted model data; 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 condition and the computing resource capacity of the federated learning participant; 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 condition and the computing resource capacity of the federated learning participant;

[0007] The application embodiment provides a data development and utilization system. The computer program is executed by the processor, so that the processor implements the data development and utilization method based on the federated learning framework.

[0008] The application embodiment provides a readable storage medium, and the readable storage medium stores programs or instructions. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the application embodiments or the related art, the following will briefly introduce the drawings needed to be used in the embodiment or related art description. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0010] Figure 1 The application embodiment provides a data development and utilization method based on a federated learning framework.

[0011] Figure 2 The application embodiment provides a data development and utilization system.

[0012] Figure 3 The application embodiment provides a data development and utilization method based on a federated learning framework.

[0013] Figure 4 The application embodiment provides a data development and utilization method based on a federated learning framework. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0015] Please refer to Figure 1 , Figure 1 is a flowchart of a data development and utilization method based on a federated learning framework provided by the embodiments of the present application. The method can be executed by a data development and utilization system, or can be executed by the data development and utilization system and a server together. The method can include steps 110-140.

[0016] Step 110: 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 locally completing model training based on non-shared original data.

[0017] In the application scenario of spicy seasonings, there are multiple federated learning participants, covering manufacturers, distributors and retailers of spicy seasonings. Each participant holds unique non-shared original data, which comes from their respective business operation links. For example, the manufacturer has relevant information of raw material procurement, including the source, quality grade, etc. of different raw materials, and production process parameters such as temperature, pressure, time, etc. The distributor masters detailed information of the sales channel, such as cooperating supermarkets, e-commerce platforms, etc., and the sales data of each channel. The retailer collects consumer feedback, such as evaluation of product taste and packaging, and market research information, such as consumer preferences and competitor product conditions, etc.

[0018] Each federated learning participant locally trains a model using its own non-shared original data. Taking the manufacturer as an example, a convolutional neural network (CNN) model in deep learning can be used, which has an input layer, a convolutional layer, a pooling layer, a fully connected layer, etc. The input layer receives raw material procurement and production process parameter data, the convolutional layer extracts features in the data through a convolution kernel, the pooling layer processes the features by dimension reduction, and the fully connected layer integrates the processed features and outputs a prediction result. In the training process, the back propagation algorithm is used to adjust the parameters of the model to minimize the error between the prediction result and the actual result.

[0019] For example, the manufacturer 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 input layer neurons is determined according to the data dimension, the raw material procurement data contains indicators such as price, quality grade, etc. of 10 different raw materials, and the production process parameters have 5 key parameters such as temperature, pressure, and time, so the number of input layer neurons is 15, and the input data is a 15-dimensional vector.

[0020] The first convolutional layer has a kernel size of 3x3 and 32 kernels. The step size is 1 and the padding method is "same" to maintain the feature map size. The activation function uses ReLU, which can effectively alleviate the gradient disappearance and make the model training more efficient. The second convolutional layer has the same kernel size of 3x3, but the number of kernels increases to 64 to further extract more complex features. The step size and padding method remain unchanged, and the activation function is still ReLU.

[0021] The pooling layer uses max pooling with a pooling window size of 2x2 and a step size of 2. This can reduce the data volume while preserving important feature information.

[0022] The first fully connected layer has 128 neurons, and the activation function uses ReLU to enhance the non-linear representation ability. The second fully connected layer is the output layer, and since it is used to predict the production efficiency, which is divided into high, medium, and low levels, the number of neurons is 3, and the Softmax activation function is used to convert the output to a probability distribution for classification.

[0023] During training, the stochastic gradient descent (SGD) optimization algorithm is used, with a learning rate of 0.001 and a momentum parameter of 0.9. The cross-entropy loss function is used to measure the difference between the predicted results and the actual results, and the training is performed for 200 rounds.

[0024] For example, the distributor local model uses a variant of the recurrent neural network (RNN), the long short-term memory network (LSTM), to process sales channel and sales data to predict sales trends. The number of input layer neurons is determined by the data dimension, and if there are 8 types of online e-commerce platforms, offline supermarkets, convenience stores, etc. in the sales channel, and the sales data contains 4 indicators such as sales in different time periods, the number of input layer neurons is 12, and the input data is a 12-dimensional vector.

[0025] The LSTM layer has 64 neurons, and the LSTM unit has a forget gate, an input gate, and an output gate, which can effectively handle the long-term dependencies of sequence data.

[0026] The fully connected layer has 1 neuron, and the activation function uses Sigmoid to convert the output to a probability value between 0 and 1, representing the probability of sales trend rising.

[0027] The Adam optimization algorithm is used for training, with a learning rate of 0.0005. The loss function is the binary cross-entropy loss function, which is suitable for binary classification problems, and the number of training rounds is 150.

[0028] For example, a retailer local model uses a multi-layer perceptron (MLP) to process consumer feedback and market research data to analyze consumer preferences. The number of input layer neurons is determined by the dimensionality of the data. If the consumer feedback includes 6 aspects such as taste, packaging, and price satisfaction, and the market research data has 4 indicators such as competitor product information and market demand trends, the number of input layer neurons is 10, and the input data is a 10-dimensional vector.

[0029] The number of neurons in the first hidden layer is 50, and the activation function is ReLU. The number of neurons in the second hidden layer is 30, and the activation function is also ReLU.

[0030] The number of neurons in the output layer is determined by the number of consumer preference categories, which are divided into 5 different preference types. The number of output layer neurons is 5, and the Softmax activation function is used to output the probability distribution.

[0031] The Adagrad optimization algorithm is used for training, with a learning rate of 0.01. The loss function is the cross-entropy loss function, and the number of training rounds is 180.

[0032] After training, each participant encrypts the model data. The encryption process uses an asymmetric encryption algorithm such as RSA. RSA is based on the difficulty of factoring large prime numbers to generate public and private keys. Participants use the public key to encrypt the model data, ensuring data security during transmission. Encrypted model data is transmitted to the data development and utilization system through a secure communication link. The SSL / TLS protocol is used for secure communication, which ensures data security and integrity during transmission through encryption and authentication mechanisms. The data development and utilization system receives these encrypted model data, verifies the integrity and authenticity of the data, such as by checking the hash value of the data to ensure that the data has not been modified during transmission.

[0033] Step 120: Perform homomorphic encryption operations on the model parameter gradient information in the encrypted model data using an asymmetric encryption algorithm, and combine the model structure metadata in the encrypted model data to obtain global model encrypted data after completing the hierarchical aggregation encryption processing.

[0034] After receiving the encrypted model data from each federated learning participant, the data development and utilization system begins processing the model parameter gradient information. Model parameter gradient information reflects the changes in model parameters during training.

[0035] In detail, the homomorphic encryption operation on the model parameter gradient information in the encrypted model data by the asymmetric encryption algorithm, and the verification combining the model structure metadata in the encrypted model data to obtain the global model encrypted data completed the hierarchical aggregation encryption processing, comprising: Step 121: Extract the model structure metadata in the encrypted model data, and parse to obtain the model hierarchical structure descriptor and parameter dimension distribution characteristics.

[0036] The data development and utilization system extracts the model structure metadata from the encrypted model data. The model structure metadata contains information such as the hierarchical structure and parameter dimensions of the model. In the parsing process, the system first identifies the key information in the metadata, such as the number of layers of the model, the type of each layer (such as convolutional layer, fully connected layer, etc.), and the number and dimensions of the parameters of each layer. Through the analysis of these information, the model hierarchical structure descriptor and parameter dimension distribution characteristics are obtained. The model hierarchical structure descriptor can be represented by a tree structure, and each node represents a model layer. The connection between nodes represents the relationship between layers. The parameter dimension distribution characteristics record the parameter dimension information of each layer, for example, for a convolutional layer, record the number, size and channel number of its convolution kernel, etc.

[0037] Step 122: Based on the model hierarchical structure descriptor, the model parameter gradient information is hierarchically divided and processed to generate gradient information subsets corresponding to the model hierarchy.

[0038] According to the model hierarchical structure descriptor obtained in step 121, the model parameter gradient information is hierarchically divided. The system assigns the model parameter gradient information to the corresponding hierarchy according to the hierarchical order of the model. For example, for a neural network model with an input layer, a hidden layer and an output layer, the parameter gradient information of the input layer is divided into the gradient information subset of the input layer, the parameter gradient information of the hidden layer is divided into the gradient information subset of the hidden layer, and the parameter gradient information of the output layer is divided into the gradient information subset of the output layer. Each gradient information subset contains all the parameter gradient information of the hierarchy.

[0039] Step 123: Homomorphic encryption operation on each gradient information subset using the public key of the asymmetric encryption algorithm to generate encrypted gradient information units.

[0040] Each gradient information subset is homomorphically encrypted using the public key of an asymmetric encryption algorithm. Homomorphic encryption allows corresponding calculations to be performed on encrypted data without decryption. Taking the Paillier homomorphic encryption algorithm as an example, the algorithm has additive homomorphism, that is, the result of performing an addition operation on two encrypted data is the same as that of performing an addition operation on the decrypted data and then encrypting it again. The system encrypts each gradient information subset using the public key to generate an encrypted gradient information unit. During encryption, the integrity and security of each encrypted gradient information unit are ensured, such as by adding an encrypted checksum to verify whether the data has been tampered with.

[0041] Step 124: Perform relevance verification on the encrypted gradient information unit and the corresponding parameter dimension distribution feature, verify the matching of the parameter dimension before and after encryption, and obtain a verified encrypted gradient set.

[0042] The generated encrypted gradient information unit is associated with the corresponding parameter dimension distribution feature. The system checks whether the parameter dimension of the encrypted gradient information unit is consistent with the parameter dimension distribution feature. If it is not consistent, it means that an error may have occurred or the data has been tampered with during the encryption process. During the verification process, the parameter dimension of each encrypted gradient information unit is checked one by one to ensure that it is consistent with the description in the parameter dimension distribution feature. If the verification is passed, these encrypted gradient information units are combined to form a verified encrypted gradient set.

[0043] Step 125: Perform hierarchical aggregation processing on the verified encrypted gradient set, reorganize the encrypted data according to the hierarchical order of the model hierarchical structure descriptor, and generate global model encrypted data that has completed hierarchical aggregation encryption processing.

[0044] The verified encrypted gradient set is subjected to hierarchical aggregation processing, and the system reorganizes the encrypted data according to the hierarchical order of the model hierarchical structure descriptor. For each level, the encrypted gradient information units of that level are aggregated, such as by integrating the encrypted gradient information units of the same level from multiple participants through weighted averaging. Then, the aggregation results of each level are combined according to the hierarchical order of the model to generate global model encrypted data that has completed hierarchical aggregation encryption processing. During the reorganization process, the hierarchical relationship and structure of the data are ensured to be consistent with the model hierarchical structure descriptor, ensuring the accuracy and effectiveness of the global model encrypted data.

[0045] Step 130: Distribute the global model encrypted data to each federated learning participant according to the preset sharding transmission strategy, which is adjusted according to the network bandwidth status and computing resource capacity of the federated learning participants.

[0046] After obtaining the global model encrypted data after the hierarchical aggregation encryption processing, the data development and utilization system needs to distribute it to each federated learning participant. Due to the different network bandwidth conditions and computing resource capacities of each participant, in order to ensure the efficiency and stability of data transmission, a preset sharding transmission strategy is adopted.

[0047] In a preferred embodiment, the global model encrypted data is distributed to each federated learning participant according to a preset sharding transmission strategy, which is adjusted according to the network bandwidth conditions and computing resource capacities of the federated learning participants, comprising: Step 131: Collect real-time network bandwidth conditions and computing resource capacity data of each federated learning participant, and generate a participant resource state vector.

[0048] The data development and utilization system collects real-time network bandwidth conditions and computing resource capacity data of each federated learning participant through communication. The network bandwidth condition reflects the network transmission capability of the participant at the current time, and the computing resource capacity data includes the memory capacity, processor core number and other information of the participant. The system integrates these data into a participant resource state vector, which contains information in multiple dimensions such as network bandwidth, memory capacity and processor core number. For example, the participant resource state vector can be represented as (network bandwidth, memory capacity, processor core number).

[0049] Step 132: Based on the participant resource state vector, a sharding transmission decision model is constructed, taking the total data volume of the global model encrypted data as an input variable and the network bandwidth condition and computing resource capacity as constraint conditions.

[0050] Based on the collected participant resource state vector, a sharding transmission decision model is constructed. The purpose of this model is to determine the optimal number of shards and shard size to meet the network bandwidth and computing resource limitations of each participant. The total data volume of the global model encrypted data is taken as an input variable, and the network bandwidth condition and computing resource capacity are taken as constraint conditions.

[0051] Optionally, the sharding transmission decision model is constructed based on the participant resource state vector, taking the total data volume of the global model encrypted data as an input variable and the network bandwidth condition and computing resource capacity as constraint conditions, comprising: Step 1321: Perform fluctuation interference optimization on the network bandwidth condition in the participant resource state vector to obtain a stable bandwidth characteristic value.

[0052] The network bandwidth status of the participants can fluctuate, and in order to obtain more accurate bandwidth information, the network bandwidth status in the participant resource state vector is optimized for fluctuation interference. The moving average method is used to smooth the network bandwidth data and remove short-term fluctuations in the data. By calculating the average value of the network bandwidth over a period of time, the stable bandwidth characteristic value is obtained, which can better reflect the actual network transmission capacity of the participants.

[0053] Step 1322: Normalize the memory capacity and processor core number in the computing resource capacity data to generate a computing resource comprehensive index.

[0054] The memory capacity and processor core number in the computing resource capacity data have different dimensions, and in order to consider them comprehensively, the memory capacity and processor core number are normalized. The minimum-maximum normalization method is used to map the memory capacity and processor core number to the [0, 1] interval. Then, according to a certain weight, the normalized memory capacity and processor core number are weighted and summed to generate a computing resource comprehensive index, which reflects the comprehensive computing capacity of the participants.

[0055] Step 1323: Construct a mathematical programming expression of the shard transmission decision model with the total data volume of the global model encrypted data as the objective function and the stable bandwidth characteristic value and the computing resource comprehensive index as the constraint conditions.

[0056] The total data volume of the global model encrypted data is used as the objective function, that is, the global model encrypted data is transmitted as efficiently as possible under the premise of meeting the network bandwidth and computing resource limitations of each participant. The stable bandwidth characteristic value and the computing resource comprehensive index are used as constraint conditions to construct a mathematical programming expression of the shard transmission decision model, which describes the relationship between the number of shards, shard size, network bandwidth, and computing resources.

[0057] Step 1324: Introduce a shard transmission delay penalty factor in the mathematical programming expression, which has a positive correlation with the number of shards and a negative correlation with the stable bandwidth characteristic value.

[0058] In order to avoid an increase in transmission delay caused by too many shards, a shard transmission delay penalty factor is introduced in the mathematical programming expression. The penalty factor has a positive correlation with the number of shards, that is, the more the number of shards, the larger the penalty factor; and has a negative correlation with the stable bandwidth characteristic value, that is, the larger the stable bandwidth characteristic value, the smaller the penalty factor. By introducing the penalty factor, the impact of transmission delay is considered in the optimization process, making the model more reasonable.

[0059] Step 1325: Solve the mathematical programming expression by the Lagrange multiplier method to obtain a feasible solution set of the optimal shard number and shard size parameters that satisfy the constraint conditions.

[0060] wherein the Lagrange multiplier method is a method for solving constrained optimization problems by introducing Lagrange multipliers to convert the constraints into part of the objective function. In the solving process, a feasible solution set of optimal number of shards and shard size parameters that satisfy the constraints is found. From these feasible solution sets, the optimal number of shards and shard size parameters are selected to ensure the efficiency and stability of data transmission.

[0061] Step 133: Calculate the optimal number of shards and shard size parameters for each federated learning participant through the shard transmission decision model.

[0062] According to the constructed shard transmission decision model, the optimal number of shards and shard size parameters for each federated learning participant are calculated. The network bandwidth conditions and computing resource capacities of each participant are different, so the optimal number of shards and shard size parameters obtained are also different. For example, participants with higher network bandwidth and more abundant computing resources may obtain larger shard size and fewer number of shards, while participants with lower network bandwidth and limited computing resources may obtain smaller shard size and more number of shards.

[0063] Step 134: Perform shard processing on the global model encrypted data according to the optimal number of shards and shard size parameters to generate an encrypted data shard set containing shard identification and sequence index.

[0064] According to the calculated optimal number of shards and shard size parameters, the global model encrypted data is processed by shards. The system divides the global model encrypted data according to the shard size to obtain multiple shards. A unique shard identification is generated for each shard, which contains information such as model layer feature code, data block type identification, and participant association marker. At the same time, a shard sequence index system is constructed, and a bidirectional linked list structure index containing predecessor shard pointer and successor shard pointer is generated by analyzing the feature dependency relationship of each shard in the global model encrypted data. The shard identification, bidirectional linked list structure index, and actual shard data are multi-dimensionally associated and packaged to form an encrypted data shard set containing feature integrity verification fields.

[0065] In an alternative embodiment, the shard processing of the global model encrypted data according to the optimal number of shards and shard size parameters to generate an encrypted data shard set containing shard identification and sequence index comprises: Step 1341: Perform feature boundary detection on the global model encrypted data according to the optimal number of shards to identify model structure feature blocks with semantic integrity in the global model encrypted data, and divide a predetermined number of logical shard units based on the natural boundaries of the model structure feature blocks.

[0066] Feature boundary detection is performed on the global model encrypted data according to the optimal number of shards. By analyzing the feature distribution in the global model encrypted data, model structure feature blocks with semantic integrity are identified. These feature blocks represent different functional parts of the model, such as different layers in a neural network model. Based on the natural boundary division of the model structure feature blocks, a pre-set number of logical shard units are divided, ensuring that each logical shard unit has a certain semantic integrity.

[0067] Step 1342: Feature-aware cutting is performed on each logical shard unit based on the shard size parameter. The feature density distribution is detected by a sliding window, and cutting operations are performed in areas where the feature density is below the density threshold, generating actual shard data.

[0068] Feature-aware cutting is performed on each logical shard unit based on the shard size parameter. The feature density distribution is detected by a sliding window method, and the sliding window slides over the logical shard unit to calculate the feature density within each window. When the feature density is below the density threshold, cutting operations are performed in that area to divide the logical shard unit into multiple actual shard data. This ensures that each actual shard data contains sufficient feature information while meeting the shard size requirements.

[0069] Step 1343: A unique shard identifier is constructed for each actual shard data, which is generated by extracting the feature fingerprint of the shard data. The shard identifier contains model level feature code, data block type identifier, and participant association marker.

[0070] A unique shard identifier is constructed for each actual shard data. The shard identifier is generated by extracting the feature fingerprint of the shard data, which is a feature representation of the shard data and has uniqueness. The shard identifier contains information such as model level feature code, data block type identifier, and 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.

[0071] Step 1344: A shard order index system is constructed, and a bidirectional linked list structure index containing predecessor shard pointers and successor shard pointers is generated by analyzing the feature dependency relationship of each actual shard data in the global model encrypted data.

[0072] A shard order index system is constructed, and the feature dependency relationship of each actual shard data in the global model encrypted data is analyzed. According to the feature dependency relationship, a bidirectional linked list structure index containing predecessor shard pointers and successor shard pointers is generated. The predecessor shard pointer points to the previous shard of the shard data, and the successor shard pointer points to the next shard of the shard data. Through the bidirectional linked list structure index, the order and dependency relationship of the shard data can be easily determined.

[0073] Step 1345: Multi-dimensionally associate and package the shard identification, doubly linked list structure index, and actual shard data, forming an encrypted data shard set containing a feature integrity verification field generated based on the feature fingerprint and sequential index relationship of the shard data.

[0074] The shard identification, doubly linked list structure index, and actual shard data are multi-dimensionally associated and packaged. By integrating these information, an encrypted data shard set containing a feature integrity verification field is formed. The feature integrity verification field is generated based on the feature fingerprint and sequential index relationship of the shard data, used to verify the integrity and correctness of the order of the shard data. During transmission, the feature integrity verification field can be checked to ensure that the shard data has not been tampered with or lost.

[0075] Step 135: Distribute the encrypted data shard set to the corresponding federated learning participants according to the sequential index of the encrypted data shard set, dynamically adjusting the transmission rate in combination with the network bandwidth status of the federated learning participants.

[0076] According to the sequential index of the encrypted data shard set, the encrypted data shard set is distributed to the corresponding federated learning participants. During transmission, the transmission rate is dynamically adjusted in combination with the network bandwidth status of the federated learning participants. If the network bandwidth of the participants is high, the transmission rate is increased; if the network bandwidth of the participants is low, the transmission rate is decreased. By dynamically adjusting the transmission rate, the efficiency and stability of data transmission are ensured. At the same time, during transmission, the transmission situation is monitored in real time, such as transmission success rate, transmission delay, etc., and problems that occur during transmission are handled in a timely manner, such as retransmitting lost shard data, etc.

[0077] By way of example, the shard transmission decision model is a mathematical programming optimization model. In terms of input variables, the total data amount of the global model encrypted data is a large scale, for example, containing 10,000 data units. The network bandwidth status of the participants 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, the 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.

[0078] When optimizing the fluctuation interference of network bandwidth status, the moving average method is set to a moving window size of 7 time units to obtain stable bandwidth characteristic values. The memory capacity and processor core number in the computing resource capacity data are normalized to generate a computing resource comprehensive index, with the memory capacity weight set to 0.65 and the processor core number weight set to 0.35.

[0079] A mathematical programming expression is constructed with the total data amount of the global model encrypted data as the objective function, and the stable bandwidth eigenvalue and the comprehensive index of computing resources as the constraint conditions. A slice transmission delay penalty factor is introduced, which has a correlation coefficient of 0.6 with the number of slices and a correlation coefficient of -0.4 with the stable bandwidth eigenvalue. The Lagrange multiplier method is used for solution, and the initial value of the Lagrange multiplier is set to 0.15, and the iteration number is 120 times.

[0080] Step 140: guiding each federated learning participant to update and optimize the local business model parameters using the global model encrypted data, and generating optimization suggestion information through model update state evaluation based on the local validation data set of each federated learning participant.

[0081] After the data development and utilization system distributes the global model encrypted data to each federated learning participant, it guides each participant to update and optimize the local business model parameters using the data. The local business model of each participant is trained based on its own non-shared original data, and the performance of the model can be further improved by combining the global model encrypted data.

[0082] In one possible design scheme, the guidance of each federated learning participant to update and optimize the local business model parameters using the global model encrypted data, and the generation of optimization suggestion information through model update state evaluation based on the local validation data set of each federated learning participant, includes: Step 141: sending a model collaborative optimization instruction to each federated learning participant, the model collaborative optimization instruction containing a homomorphic decryption rule of the global model encrypted data and a parameter feature association protocol.

[0083] The data development and utilization system sends a model collaborative optimization instruction to each federated learning participant, which contains a homomorphic decryption rule of the global model encrypted data and a parameter feature association protocol. The homomorphic decryption rule is used to guide the participant to decrypt the global model encrypted data, and 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 that the features are matched according to the hierarchical structure of the model, or the features are associated according to the semantic information of the features.

[0084] Step 142: 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 containing model hierarchical features and gradient evolution features shared across participants.

[0085] After receiving the model collaborative optimization instruction, each federated learning participant decrypts the global model encrypted data according to the homomorphic decryption rule. Then, the global parameter feature set is extracted from the decrypted global model data. The global parameter feature set includes model layer feature and gradient evolution feature shared across participants. The model layer feature reflects the layer structure and parameter distribution of the model, and the gradient evolution feature reflects the parameter change of the model in the training process.

[0086] Preferably, the global parameter feature set extracted by each federated learning participant based on the decrypted global model encrypted data comprises: Step 1421: Make the federated learning participant perform homomorphic decryption operation on the global model encrypted data through the preset decryption protocol, and separate the hierarchical parameter feature sequence and the gradient evolution trajectory feature of the global model.

[0087] The federated learning participant performs homomorphic decryption operation on the global model encrypted data through the preset decryption protocol. Homomorphic decryption operation allows encrypted data to be decrypted without revealing data privacy. Through the decryption operation, the hierarchical parameter feature sequence and the gradient evolution trajectory feature of the global model are separated. The hierarchical parameter feature sequence records the parameter values of each layer of the model, and the gradient evolution trajectory feature records the change trajectory of the model parameters in the training process.

[0088] Step 1422: Make the federated learning participant perform feature discretization processing on the hierarchical parameter feature sequence, and map the continuous parameter space to a feature symbol sequence.

[0089] The hierarchical parameter feature sequence is processed by feature discretization. Since the parameter values in the hierarchical parameter feature sequence are continuous, in order to facilitate subsequent processing and analysis, the continuous parameter space is mapped to a feature symbol sequence. A clustering algorithm is used to divide the continuous parameter values into different categories, and each category is represented by a feature symbol. For example, the K-means clustering algorithm is used to divide the parameter values into K categories, and each category corresponds to a feature symbol.

[0090] Step 1423: Make the federated learning participant perform time series feature extraction on the gradient evolution trajectory feature, identify the direction correlation and amplitude correlation of gradient change, and construct a gradient feature correlation graph.

[0091] The time series feature extraction is performed on the gradient evolution trajectory feature, and the direction correlation and amplitude correlation of gradient change are identified by analyzing the time series of the gradient evolution trajectory feature. The direction correlation reflects whether the direction of gradient change is consistent, and the amplitude correlation reflects the amplitude size relationship of gradient change. Based on these correlations, a gradient feature correlation graph is constructed. The gradient feature correlation graph uses the structure of the graph to represent the correlation between gradients, and the node represents the gradient and the edge represents the correlation between gradients.

[0092] Step 1424: instructing the federated learning participants to perform feature fusion on the feature symbol sequence and the gradient feature association graph to generate a global parameter feature set containing hierarchical structure features and evolution dynamic features.

[0093] The feature fusion on the feature symbol sequence and the gradient feature association graph integrates the information of both to generate a global parameter feature set containing hierarchical structure features and evolution dynamic features. The hierarchical structure features are represented by the feature symbol sequence, and the evolution dynamic features are represented by the gradient feature association graph. Feature fusion can be done by concatenating the information of the feature symbol sequence and the gradient feature association graph to form a new feature set.

[0094] Step 1425: receiving the global parameter feature set under the premise that it contains all parameter feature components contributed by the federated learning participants.

[0095] The data development and utilization system receives the global parameter feature set under the premise that it contains all parameter feature components contributed by the federated learning participants. Ensuring the integrity of the global parameter feature set for subsequent accurate analysis and processing. If the parameter feature components of a participant are missing, it may affect the performance and optimization effect of the model.

[0096] Step 143: instructing the federated learning participants to perform feature association matching between the global parameter feature set and the current parameter feature set of the local business model to identify complementary areas and conflict areas in the feature space.

[0097] The data development and utilization system instructs the federated learning participants to perform feature association matching between the global parameter feature set and the current parameter feature set of the local business model. Through matching, complementary areas and conflict areas in the feature space are identified. Complementary areas represent parts that can be mutually complementary between the global parameter feature set and the local parameter feature set, and conflict areas represent parts with contradictions or repetitions between the two.

[0098] Among them, the instruction to the federated learning participants to perform feature association matching between the global parameter feature set and the current parameter feature set of the local business model to identify complementary areas and conflict areas in the feature space includes instructing the federated learning participants to perform the following operations: Step 1431: extracting the current parameter feature set of the local business model, including the local hierarchical feature symbol sequence and the local gradient association graph.

[0099] The participant extracts a 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 performing feature discretization processing on the hierarchical parameters of the local business model, and the local gradient correlation graph is obtained by performing feature extraction and correlation analysis on the gradient evolution track of the local business model.

[0100] In the embodiment of the application, the feature discretization processing uses a K-means clustering algorithm, the value of K is set to 6, and the continuous parameter space is divided into 6 categories. The maximum number of iterations is set to 60 times, and the initial clustering center is randomly selected.

[0101] Step 1432: Perform feature space alignment processing on the global parameter feature set and the local parameter feature set to unify the coding rules of the feature symbol sequence and the topological structure of the gradient correlation graph.

[0102] Since the global parameter feature set and the local parameter feature set may use different coding rules and topological structures, in order to accurately match, the coding rules of the feature symbol sequence and the topological structure of the gradient correlation graph need to be unified. For example, the feature symbol sequences of the global parameter feature set and the local parameter feature set are uniformly coded in the same way, and the topological structure of the gradient correlation graph is adjusted to have consistency.

[0103] Step 1433: Perform hierarchical comparison on the aligned feature set by a feature similarity calculation method to generate a feature similarity matrix, and the matrix elements represent the correlation degree of the corresponding hierarchical features.

[0104] The feature similarity calculation method is used to compare the aligned feature set at each hierarchical level, and the cosine similarity method is used to calculate the similarity of the aligned feature set at each hierarchical level. The calculated similarity values form a feature similarity matrix, and the matrix elements represent the correlation degree of the corresponding hierarchical features. The higher the similarity value, the higher the correlation degree of the corresponding hierarchical features.

[0105] Step 1434: Based on the feature similarity matrix, identify the feature region with a similarity higher than a preset threshold as a complementary region, and the feature region with a similarity lower than the preset threshold as a conflict region.

[0106] Based on the feature similarity matrix, the complementary region and the conflict region are identified, a preset threshold is set, the feature region with a similarity higher than the preset threshold is regarded as the complementary region, and the feature region with a similarity lower than the preset threshold is regarded as the conflict region. The selection of the preset threshold can be adjusted according to the actual situation to balance the division of the complementary region and the conflict region.

[0107] Step 1435: Mark the feature boundaries of the complementary region and the conflict region, and generate a feature region division result containing region position identification and feature association type.

[0108] Mark the feature boundaries of the complementary region and the conflict region. By marking the feature boundaries, the scope of the complementary region and the conflict region is clear. Generate a feature region division result containing region position identification and feature association type. The region position identification represents the position of the complementary region and the conflict region in the feature space, and the feature association type represents whether the region is a complementary region or a conflict region.

[0109] Step 144: Adjust the hierarchical association degree of the network structure of the local business model based on the feature association matching result, which includes strengthening the cross-hierarchical connection weight through the feature complementary region and suppressing the redundant hierarchical node through the conflict region.

[0110] Adjust the hierarchical association degree of the network structure of the local business model based on the feature association matching result. For the feature complementary region, strengthen the cross-hierarchical connection weight to enhance the model's ability to utilize complementary information. For example, in a neural network model, increase the connection weight between the layers corresponding to the complementary region. For the conflict region, suppress the redundant hierarchical node to reduce the complexity and computational load of the model. For example, delete the redundant node corresponding to the conflict region or reduce its connection weight.

[0111] Step 145: Obtain the feature response sequence obtained by each federated learning participant using the local validation dataset to test the adjusted model, and generate optimization suggestion information containing parameter feature adaptation suggestions and structure association optimization directions based on the consistency and volatility analysis of the feature response sequence.

[0112] Each federated learning participant uses the local validation dataset to test the adjusted model. By inputting the local validation dataset, the output of each level of the model is recorded to obtain the feature response sequence. Based on the consistency and volatility analysis of the feature response sequence, generate optimization suggestion information containing parameter feature adaptation suggestions and structure association optimization directions.

[0113] Illustratively, the obtaining of the feature response sequence obtained by each federated learning participant using the local validation dataset to test the adjusted model, and the generation of optimization suggestion information containing parameter feature adaptation suggestions and structure association optimization directions based on the consistency and volatility analysis of the feature response sequence, include: Step 1451: instruct the federal learning participant to divide the local verification dataset into a basic feature subset, an edge feature subset, and an abnormal feature subset according to feature types, input the adjusted local business model in turn, collect the hierarchical feature response subsequence and the cross-hierarchical associated 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.

[0114] The federal learning participant is instructed to divide the local verification dataset into a basic feature subset, an edge feature subset, and an abnormal feature subset according to feature types. The basic feature subset contains the main feature information in the dataset, the edge feature subset contains some secondary and edge feature information, and the abnormal feature subset contains abnormal data in the dataset. These subsets are input into the adjusted local business model in turn, and the hierarchical feature response subsequence and the cross-hierarchical associated feature response subsequence output by each level of the model are collected. The hierarchical feature response subsequence reflects the output of the model at each level, and the cross-hierarchical associated feature response subsequence reflects the association between different levels. These subsequences are integrated to form a feature response sequence containing intra-level response and inter-level association.

[0115] Step 1452: perform multi-scale time series pattern analysis on the feature response sequence, identify the pattern evolution trend of the hierarchical feature response subsequence corresponding to the continuous input sample within the first time window, and detect the periodic association strength change of the cross-hierarchical associated feature response subsequence within the second time window, to generate a time series pattern stability map.

[0116] Perform multi-scale time series pattern analysis on the feature response sequence. Within the first time window, analyze the pattern evolution trend of the hierarchical feature response subsequence corresponding to the continuous input sample. By observing the changes in the hierarchical feature response subsequence, identify the evolution direction and trend of its pattern. Within the second time window, detect the periodic association strength change of the cross-hierarchical associated feature response subsequence. By analyzing the periodicity of the cross-hierarchical associated feature response subsequence, determine the change of its association strength. Based on these analysis results, generate a time series pattern stability map. The time series pattern stability map represents the time series pattern stability of the feature response sequence in graphical form, from which potential problems and optimization directions can be found.

[0117] Step 1453: perform spatial conduction path analysis on the feature response sequence, construct a feature conduction directed graph based on the activation order of the hierarchical feature response subsequence, identify feature blocking nodes and redundant conduction branches in the conduction path through graph structure analysis, and generate a spatial conduction efficiency evaluation matrix.

[0118] A 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 subsequences. The nodes in the graph represent the layers of the model, and the edges represent the feature conduction relationship between the layers. Through graph structure analysis, feature blocking nodes and redundant conduction branches in the conduction path are identified. Feature blocking nodes refer to nodes that hinder the transmission of features during the conduction process, and redundant conduction branches refer to branches that do not make substantial contributions 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 the form of a matrix, based on which problems in the spatial conduction path can be identified.

[0119] Step 1454: Combine the temporal pattern stability atlas and the spatial conduction efficiency evaluation matrix to construct a multi-dimensional feature response quality evaluation model, which includes a temporal consistency dimension, a spatial conduction dimension, and a feature robustness dimension. The feature robustness dimension is determined by analyzing the feature response decay rate and recovery ability of the abnormal feature subset.

[0120] A multi-dimensional feature response quality evaluation model is constructed by combining the temporal pattern stability atlas and the spatial conduction efficiency evaluation matrix. This model includes a temporal consistency dimension, a spatial conduction dimension, and a feature robustness dimension. The temporal consistency dimension reflects the consistency of the feature response sequence over time, the spatial conduction dimension reflects the conduction efficiency of features in the model, and the feature robustness dimension is determined by analyzing the feature response decay rate and recovery ability of the abnormal feature subset. The feature response decay rate of the abnormal feature subset reflects the degree of performance decline of the model when encountering abnormal data, and the recovery ability reflects the ability of the model to recover normal performance after the influence of abnormal data disappears.

[0121] Step 1455: Output the feature response quality index of each dimension through the multi-dimensional feature response quality evaluation model, and identify the temporal fragile region with a temporal consistency index below the baseline value, the conduction blocking region with a spatial conduction efficiency index below the baseline value, and the anti-interference weak region with a robustness index below the baseline value.

[0122] The feature response quality index of each dimension is output through the multi-dimensional feature response quality evaluation model. The temporal consistency index reflects the consistency of the feature response sequence over time, the spatial conduction efficiency index reflects the conduction efficiency of features in the model, and the robustness index reflects the anti-interference ability of the model. The temporal fragile region with a temporal consistency index below the baseline value, the conduction blocking region with a spatial conduction efficiency index below the baseline value, and the anti-interference weak region with a robustness index below the baseline value are identified. The baseline value can be set according to actual conditions to determine the performance standard of the model in each dimension.

[0123] Step 1456: For the timing vulnerable area, generate parameter feature timing alignment suggestions based on pattern evolution trend analysis, including feature response delay compensation mechanism and cross-window pattern association weight adjustment strategy; for the conduction block area, generate structure association optimization direction based on feature conduction directed graph path analysis, including adjacent level connection strengthening of block nodes and node pruning scheme of redundant branches; for the anti-interference weak area, generate feature enhancement adaptation suggestions combined with the response attenuation characteristics of abnormal feature subsets, including the replacement of activation function types of abnormal feature sensitive levels and the expansion strategy of feature mapping space.

[0124] For the timing vulnerable area, generate parameter feature timing alignment suggestions based on pattern evolution trend analysis, including feature response delay compensation mechanism and cross-window pattern association weight adjustment strategy. Feature response delay compensation mechanism is used to compensate for the delay of feature response in time, and cross-window pattern association weight adjustment strategy is used to adjust the pattern association weight between different time windows to improve the timing consistency. For the conduction block area, generate structure association optimization direction based on feature conduction directed graph path analysis, including adjacent level connection strengthening of block nodes and node pruning scheme of redundant branches. By strengthening the adjacent level connection of block nodes and pruning the nodes of redundant branches, the spatial conduction efficiency is improved. For the anti-interference weak area, generate feature enhancement adaptation suggestions combined with the response attenuation characteristics of abnormal feature subsets, including the replacement of activation function types of abnormal feature sensitive levels and the expansion strategy of feature mapping space. By replacing the activation function types of abnormal feature sensitive levels and expanding the feature mapping space, the anti-interference ability of the model is improved.

[0125] In the feature conduction directed graph construction, the graph nodes represent the model levels, and if the model has 6 levels, there are 6 nodes. The edge represents the feature conduction relationship between the levels, and the edge weight is determined by the activation strength of the feature response subsequence of the level. The greater the activation strength, the greater the weight. Feature block nodes and redundant conduction branches are identified through graph structure analysis.

[0126] Step 1457: Based on the timing alignment suggestions, structure association optimization direction and feature enhancement adaptation suggestions, generate optimization suggestion information containing multi-dimensional optimization priority sorting.

[0127] Based on the timing alignment suggestions, structure association optimization direction and feature enhancement adaptation suggestions, generate optimization suggestion information containing multi-dimensional optimization priority sorting. According to the importance and urgency of each suggestion for model performance improvement, the priority of these suggestions is sorted. Multi-dimensional optimization priority sorting considers factors such as timing consistency, spatial conduction and feature robustness, to ensure that the model can be effectively optimized in multiple aspects.

[0128] In the embodiments of the present application, the multi-dimensional feature response quality evaluation model uses a multi-layer perception (MLP) to evaluate the feature response quality. The number of input layer neurons is determined according to the dimension of the feature response sequence. If there are 12 dimensional features including both intra-level responses and inter-level correlations, the number of input layer neurons is 12, and the input data is a 12-dimensional vector.

[0129] The number of neurons in the first hidden layer is 30, and the activation function is ReLU. The number of neurons in the second hidden layer is 20, and the activation function is ReLU. The number of neurons in the output layer is 3, corresponding to the feature response quality indexes of the temporal consistency dimension, the spatial conductivity dimension, and the feature robustness dimension, respectively, and the activation function is a linear function directly outputting the indexes. The Adagrad optimization algorithm is used for training, and the learning rate is 0.008. The mean square error loss function is used as the loss function, and the training rounds are 220 rounds.

[0130] In an extended embodiment, the method further comprises: Step 210: Before uploading the encrypted model data on each federal learning participant, guide the federal learning participant to perform feature desensitization preprocessing on the local non-shared original data, remove the identity features and sensitive attribute features in the original data through feature abstraction processing, and retain the structural features and correlation features required for model training.

[0131] Before uploading the encrypted model data on each federal learning participant, guide it to perform feature desensitization preprocessing on the local non-shared original data. The purpose of feature desensitization preprocessing is to protect the privacy and security of data while retaining the effective information required for model training. Through feature abstraction processing, identity features and sensitive attribute features in the original data are removed. Identity features include user's name, ID number, etc., and sensitive attribute features include user's health information, financial information, etc. The structural features and correlation features required for model training are retained, such as the correlation between raw material procurement and production process parameters, the correlation between sales channels and sales volume, etc. Feature abstraction processing can use generalization, masking, etc. Generalization method replaces specific feature values with more abstract categories, such as replacing specific age values with age intervals; masking method replaces sensitive feature values with masking symbols, such as replacing part of the digits of ID number with asterisks.

[0132] Step 220: After receiving the desensitized encrypted model data, perform reversibility verification on the feature abstraction results in the data to ensure that the desensitization process does not destroy the key feature correlation required for model training, and perform hierarchical aggregation encryption processing after verification.

[0133] After receiving the desensitized encrypted model data, the feature abstraction results in the data are verified for reversibility. The purpose of reversibility verification is to ensure that the desensitization process does not destroy the key feature correlation required for model training. By performing the reverse operation on the feature abstraction results, it is checked whether the key feature correlation of the original data can be restored. For example, for the features after generalization processing, it is checked whether the approximate range of the original features can be restored according to the generalization rules; for the features after mask processing, it is checked whether the key feature correlation can be preserved without leaking sensitive information. If the verification is passed, perform hierarchical aggregation encryption processing, as described in step 120.

[0134] Step 230: In the global model encrypted data distribution stage, generate a dedicated feature mask for each federated learning participant, which is dynamically generated based on the participant's desensitized feature distribution, and is used to mask the desensitized features of other federated learning participants irrelevant to the federated learning participant.

[0135] In the global model encrypted data distribution stage, a dedicated feature mask is generated for each federated learning participant. The dedicated feature mask is dynamically generated based on the participant's desensitized feature distribution. By analyzing the participant's desensitized feature distribution, it is determined which features are related to the participant and which features are irrelevant to the participant. The dedicated feature mask is used to mask the desensitized features of other federated learning participants irrelevant to the participant, ensuring that the participant can only obtain feature information related to its own business. For example, for a manufacturer, the dedicated feature mask can mask the sensitive feature information of distributors and retailers, and only retain the feature information related to production.

[0136] Step 240: After the federated learning participant receives the global model encrypted data containing the dedicated feature mask, filter out irrelevant features through the mask and retain feature components related to the local business model for parameter updating and structure optimization.

[0137] After the federated learning participant receives the global model encrypted data containing the dedicated feature mask, filter out irrelevant features through the mask. The participant uses the dedicated feature mask to filter the global model encrypted data, and only retains feature components related to the local business model. Then, use these retained feature components to update and optimize the parameters of the local business model, as described in step 140. By filtering out irrelevant features through the mask, it is ensured that the participant only uses information related to its own business when updating and optimizing the model, improving the relevance and accuracy of the model.

[0138] Step 250: Test whether sensitive features are effectively isolated by simulating an attack, jointly evaluate the feature desensitization effect and the mask filtering effect, and adjust the feature desensitization rules and the mask generation strategy to maintain a dynamic balance between privacy protection and model performance according to the evaluation results.

[0139] The sensitive features are verified to be effectively isolated through simulation attack testing. The simulation attack testing adopts various attack means such as data mining, machine learning attack, etc., to attempt to recover the sensitive feature information from the desensitized encrypted model data and the filtered feature components. If the sensitive feature information cannot be recovered in the simulation attack testing, it indicates that the sensitive features are effectively isolated. The effect of feature desensitization and the effect of mask filtering are jointly evaluated. The evaluation indexes include the degree of privacy protection and the degree of model performance loss. The degree of privacy protection reflects the degree of protection of the sensitive features, and the degree of model performance loss reflects the influence of feature desensitization and mask filtering on the model performance. According to the evaluation results, the feature desensitization rule and the mask generation strategy are adjusted. If the degree of privacy protection is not enough, the feature desensitization rule and the mask generation strategy are strengthened; if the model performance loss is too large, the feature desensitization rule and the mask generation strategy are appropriately relaxed to maintain the dynamic balance between privacy protection and model performance.

[0140] In an extended embodiment, the method further comprises: Step 310: Constructing a federated learning feature interaction behavior feature sequence library, recording the feature interaction behavior feature sequences of each federated learning participant in the process of uploading, downloading and parameter updating of encrypted model data, the behavior feature sequence containing interaction type identification, model layer level associated feature, feature flow direction vector and interaction context associated feature.

[0141] A federated learning feature interaction behavior feature sequence library is constructed, which is used to record the feature interaction behavior feature sequences of each federated learning participant in the process of uploading, downloading and parameter updating of encrypted model data. The behavior feature sequence contains interaction type identification, model layer level associated feature, feature flow direction vector and interaction context associated feature. The interaction type identification represents the type of interaction, such as uploading, downloading, parameter updating, etc.; the model layer level associated feature represents the model layer involved in the interaction; the feature flow direction vector represents the flow direction of the feature; and the interaction context associated feature represents the context information of the interaction, such as time, place, etc. By recording these feature interaction behavior feature sequences, the behavior of the participants can be analyzed and monitored, and abnormal behavior can be found in time.

[0142] Step 320: Extracting context associated features from the feature interaction behavior feature sequence through a time sequence association algorithm, establishing a cross-participant feature interaction behavior association graph, and identifying feature interaction paths with abnormal connection strength in the graph.

[0143] Contextual correlation features of the feature interaction behavior feature sequence are extracted by a time sequence correlation algorithm. The time sequence correlation algorithm can analyze the time sequence and correlation of the feature interaction behavior feature sequence and extract contextual correlation features therefrom. Based on these contextual correlation features, a cross-participant feature interaction behavior correlation graph is established. The nodes in the graph represent feature interaction behaviors of participants, and the edges represent correlation between the feature interaction behaviors. By analyzing the connection strength of the edges in the graph, a feature interaction path with abnormal connection strength is identified. The abnormal connection strength may indicate abnormal feature interaction behaviors, such as data leakage, malicious attacks, etc.

[0144] Step 330: input the feature interaction behavior feature sequence into a pre-trained time sequence anomaly detection model, and perform weighted processing on the interaction context correlation features by a feature attention mechanism module of the model to generate an abnormal behavior probability distribution of the feature interaction behavior.

[0145] The feature interaction behavior feature sequence is input into a pre-trained time sequence anomaly detection model. The model uses structures such as recurrent neural networks (RNN) or long short-term memory networks (LSTM) in deep learning, and has time sequence processing capability. The feature attention mechanism module of the model performs weighted processing on the interaction context correlation features, highlighting important context correlation features. Through the processing of the model, an abnormal behavior probability distribution of the feature interaction behavior is generated, which represents the probability that each feature interaction behavior is an abnormal behavior.

[0146] Step 340: based on the abnormal behavior probability distribution, filter out feature interaction behavior segments with an abnormal probability value exceeding a set probability value, and perform feature variation pattern analysis on the model level correlation features and feature flow direction vectors in the feature interaction behavior segments to obtain abnormal behavior segments.

[0147] Based on the abnormal behavior probability distribution, feature interaction behavior segments with an abnormal probability value exceeding a set probability value are filtered out. A probability threshold is set, and feature interaction behavior segments with an abnormal probability value exceeding the threshold are filtered out. Feature variation pattern analysis is performed on the model level correlation features and feature flow direction vectors in these abnormal behavior segments. Feature variation pattern analysis can find the variation patterns of features in abnormal behavior segments, such as sudden changes in feature values, abnormal feature flow directions, etc. Based on the analysis results, abnormal behavior segments are obtained, and the specific features and forms of abnormal behaviors are determined.

[0148] Step 350: combine feature gene fingerprint extraction, interaction path topology structure analysis, and participant behavior baseline comparison to determine the risk level and potential threat type of the abnormal behavior segments, and obtain a risk assessment result; generate a dynamic response strategy based on the risk level and potential threat type, the dynamic response strategy including a feature isolation strategy, an interaction flow limiting strategy, and a path redirection strategy.

[0149] In combination with feature gene fingerprint extraction, interactive path topology analysis, and participant behavior baseline comparison, the risk level and potential threat type of the abnormal behavior segment are determined. Feature gene fingerprint extraction can identify the unique characteristics of the abnormal behavior segment, interactive path topology analysis can analyze the propagation path of the abnormal behavior, and participant behavior baseline comparison can compare the differences between the abnormal behavior and the normal behavior of the participant. Based on these analysis results, the risk level and potential threat type of the abnormal behavior segment are determined, such as low risk, medium risk, high risk, data leakage, malicious attack, etc. A dynamic response strategy is generated according to the risk level and potential threat type. The dynamic response strategy includes feature isolation strategy, interactive flow limiting strategy, and path redirection strategy. The feature isolation strategy is used to isolate the features involved in the abnormal behavior to prevent its propagation; the interactive flow limiting strategy is used to limit the interactive flow of the abnormal behavior to reduce its impact range; the path redirection strategy is used to change the interactive path of the abnormal behavior to guide it away from sensitive data and critical systems.

[0150] For example, feature gene fingerprint extraction uses the MD5 hash algorithm to generate a 128-bit hash value as a feature gene fingerprint for the abnormal behavior segment. When determining the risk level and potential threat type of the abnormal behavior, similar abnormal behaviors are identified by comparing fingerprints.

[0151] Step 360: The abnormal behavior segment, risk assessment result, and dynamic response strategy execution effect are fed back to the feature interaction behavior feature sequence library as security knowledge metadata, and the detection parameters of the time series anomaly detection model and the feature weight coefficients of the multi-dimensional feature traceability analysis are updated.

[0152] The abnormal behavior segment, risk assessment result, and dynamic response strategy execution effect are fed back to the feature interaction behavior feature sequence library as security knowledge metadata. These security knowledge metadata can be used to update the detection parameters of the time series anomaly detection model and the feature weight coefficients of the multi-dimensional feature traceability 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 effect of the multi-dimensional feature traceability analysis is optimized to better discover and handle abnormal behaviors. At the same time, security knowledge metadata is continuously accumulated to improve the security and reliability of the system.

[0153] In the embodiments of the present application, the time series anomaly detection model uses a long short-term memory network (LSTM) to monitor feature interaction behavior. The number of input layer neurons is determined according to the feature interaction behavior feature sequence dimension. If it contains 4-dimensional information such as interactive type identification, model level correlation feature, feature flow direction vector, and interactive context correlation feature, the number of input layer neurons is 4, and the input data is a 4-dimensional vector.

[0154] The number of LSTM layer neurons is set to 48, and the long-term dependence of sequence data is processed through the forgetting gate, input gate and output gate. The number of neurons in the fully connected layer is 1, and the activation function uses Sigmoid to output the probability of abnormal behavior. The training uses the Adam optimization algorithm, and the learning rate is 0.0002. The loss function uses the binary cross-entropy loss function, and the training rounds are 250 rounds.

[0155] Referring to Figure 2 The figure is a schematic diagram of the basic structure of a data development and utilization system 200 provided by the embodiments of the present application, which comprises a processor 201, a storage device 202 having a computer program 2020 stored thereon, 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.

[0156] Referring to Figure 3 The functional module block diagram of the data development and utilization device based on the federated learning framework provided by the embodiments of the present application, the data development and utilization device based on the federated learning framework comprises: A data receiving module is configured to receive encrypted model data transmitted by each federated learning participant through a secure communication link, wherein the encrypted model data is generated by each federated learning participant after completing model training based on non-shared original data locally; A data encryption module is configured to perform homomorphic encryption operation on model parameter gradient information in the encrypted model data through an asymmetric encryption algorithm, and obtain global model encrypted data after completing hierarchical aggregation encryption processing by combining and verifying model structure metadata in the encrypted model data; A data distribution module is configured to distribute the global model encrypted data to each federated learning participant according to a preset sharding transmission strategy, wherein the sharding transmission strategy is adjusted according to the network bandwidth status and the computing resource capacity of the federated learning participant; A federated training module is configured to guide each federated learning participant to update parameters and optimize structure of a local business model using the global model encrypted data, and generate optimization suggestion information by evaluating the model update state through a local validation data set of each federated learning participant.

[0157] On the basis of the above, a readable storage medium is provided, the readable storage medium has a program or instructions stored thereon, and the program or instructions are executed by a processor to implement the steps of the above method.

[0158] In summary, please refer to Figure 4From the perspective of data utilization, the embodiments of the present application break through the barriers of scattered data and unwillingness to share in the production, sales, and marketing links by receiving encrypted model data generated by each federal learning participant based on non-shared original data, so that each party can mine more value from multi-party data while protecting data security and privacy. For example, after combining the data in the production and sales links, more accurate production and sales decision-making basis can be provided for enterprises.

[0159] In terms of model processing, the embodiments of the present application perform homomorphic encryption operation on the model parameter gradient information in the encrypted model data and combine model structure metadata verification to complete hierarchical aggregation encryption processing to obtain global model encrypted data, which not only ensures the security of data in the aggregation process, but also effectively integrates the model advantages of each party to improve the performance and accuracy of the global model.

[0160] In terms of data transmission, the embodiments of the present application adjust the fragmentation transmission strategy according to the network bandwidth status and computing resource capacity of the federal learning participants to ensure that the global model encrypted data can be efficiently and stably distributed to each participant, avoiding data transmission bottlenecks caused by network and computing resource differences and improving the overall efficiency and adaptability of the system.

[0161] In the model optimization link, the embodiments of the present application guide each participant to update and optimize the local business model using the global model encrypted data, and evaluate the optimization suggestion information through the local validation data set, so that the local business model of each participant can be continuously optimized, realizing the whole-process optimization from data to model to business.

[0162] In addition, it should be noted that: the embodiments of the present application also provide a computer program product, which can 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 to make the data development and utilization system execute the foregoing Figure 1 The description of the method in the corresponding embodiment, therefore, will not be repeated here. In addition, the beneficial effects of using the same method will not be repeated. For technical details not disclosed in the computer program product embodiments involved in the present application, please refer to the description of the method embodiments of the present application.

[0163] It should be noted that the embodiments in the present specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system or device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant part can be referred to the method part.

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 hierarchical aggregation encryption processing through verification combining model structure metadata in the encrypted model data; 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 using the global model encrypted data, and generating optimization suggestion information through model updating state evaluation by a local validation data set of each federated learning participant.

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

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

4. The method of claim 3, wherein, The slice transmission decision model is constructed based on the participant resource state vector, the total data amount of the global model encrypted data is taken as an input variable, the network bandwidth condition and the computing resource capacity are taken as constraint conditions, and the slice transmission decision model includes: The network bandwidth condition in the participant resource state vector is subjected to fluctuation interference optimization to obtain a stable bandwidth characteristic value; The memory capacity and the number of processor cores in the computing resource capacity data are subjected to normalization processing to generate a computing resource comprehensive index; A mathematical programming expression of the slice transmission decision model is constructed with the total data amount of the global model encrypted data as a target function and the stable bandwidth characteristic value and the computing resource comprehensive index as constraint conditions; A slice transmission delay penalty factor is introduced in the mathematical programming expression, the delay penalty factor has a positive correlation with the number of slices and a negative correlation with the stable bandwidth characteristic value; The mathematical programming expression is solved by the Lagrange multiplier method to obtain a feasible solution set of the optimal number of slices and slice size parameters that meet the constraint conditions.

5. The method of claim 4, wherein, The global model encrypted data is sliced according to the optimal number of slices and slice size parameters to generate an encrypted data slice set containing slice identification and order index, and the method includes: Feature boundary detection is performed on the global model encrypted data according to the optimal number of slices to identify model structure feature blocks with semantic integrity in the global model encrypted data, and a preset number of logical slice units are divided based on the natural boundaries of the model structure feature blocks; Each logical slice unit is subjected to feature-aware cutting based on the slice size parameter, feature density distribution is detected through a sliding window, and cutting operation is performed in a region where the feature density is lower than a density threshold to generate actual slice data; A unique slice identification is constructed for each actual slice data, the slice identification is generated by extracting a feature fingerprint of the slice data, and the slice identification contains a model hierarchical feature code, a data block type identification and a participant association marker; A slice order index system is constructed, a bidirectional linked list structure index containing a predecessor slice pointer and a successor slice pointer is generated by analyzing the feature dependency relationship of each actual slice data in the global model encrypted data; The slice identification, bidirectional linked list structure index and actual slice data are multi-dimensionally associated and encapsulated to form an encrypted data slice set containing a feature integrity check field, and the feature integrity check field is generated based on the feature fingerprint and order index relationship of the slice data.

6. The method of claim 1, wherein, The global model encrypted data is used to guide each federated learning participant to perform parameter updating and structure optimization on a local business model, and an optimization suggestion information is generated by performing model updating state evaluation on a local verification data set of each federated learning participant, and the method includes: A model collaborative optimization instruction is sent to each federated learning participant, and the model collaborative optimization instruction contains a homomorphic decryption rule of the global model encrypted data and a parameter feature association protocol. receive a global parameter feature set extracted by each federal learning participant based on the decrypted global model encrypted data, the global parameter feature set containing model level features and gradient evolution features shared across participants; guide each federal learning participant to perform feature correlation matching between the global parameter feature set and a current parameter feature set of a local business model, and identify complementary areas and conflict areas in the feature space; perform level correlation degree adjustment on the network structure of the local business model based on the feature correlation matching result, the level correlation degree adjustment including strengthening cross-level connection weights through the complementary areas and suppressing redundant level nodes through the conflict areas; obtain a feature response sequence obtained by each federal learning participant through feature response testing on the adjusted model using a local verification data set, and generate optimization suggestion information including parameter feature adaptation suggestions and structure correlation optimization directions based on consistency and volatility analysis of the feature response sequence.

7. The method of claim 6, wherein, The receiving of the global parameter feature set extracted by each federal learning participant based on the decrypted global model encrypted data includes: making the federal learning participant perform homomorphic decryption operation on the global model encrypted data through a preset decryption protocol to separate a level parameter feature sequence and a gradient evolution trajectory feature of the global model; making the federal learning participant perform feature discretization processing on the level parameter feature sequence to map a continuous parameter space into a feature symbol sequence; making the federal learning participant perform time sequence feature extraction on the gradient evolution trajectory feature to identify direction correlation and amplitude correlation of gradient changes and construct a gradient feature correlation graph; making the federal learning participant perform feature fusion on the feature symbol sequence and the gradient feature correlation graph to generate the global parameter feature set containing level 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 federal learning participants; The guiding of each federal learning participant to perform feature correlation matching between the global parameter feature set and a current parameter feature set of a local business model, and identifying complementary areas and conflict areas in the feature space includes: making the federal learning participant perform the following operations: extracting a current parameter feature set of the local business model, including a local level 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 encoding rules of the feature symbol sequence and topological structures of the gradient correlation graph; performing layer-by-level comparison on the aligned feature sets through a feature similarity calculation method to generate a feature similarity matrix, and the matrix elements represent correlation degrees of corresponding level features; identifying, based on the feature similarity matrix, feature areas with similarity higher than a preset threshold as complementary areas and feature areas with similarity lower than the preset threshold as conflict areas; performing feature boundary marking on the complementary areas and the conflict areas to generate feature area division results containing area position identification and feature correlation types.

8. The method of claim 6, wherein, The acquisition of each federal learning participant uses the local verification dataset to obtain the feature response sequence of the adjusted model, and generates optimization suggestion information containing parameter feature adaptation suggestion and structure correlation optimization direction based on the consistency and volatility analysis of the feature response sequence, including: Indicate the federal learning participant to divide the local verification dataset into a basic feature subset, an edge feature subset, and an abnormal feature subset according to the feature type, input the adjusted local business model in turn, collect the hierarchical feature response subsequence and the cross-hierarchical correlation 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 correlation; Perform multi-scale time series pattern analysis on the feature response sequence, 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 correlation strength change of the cross-hierarchical correlation feature response subsequence in the second time window, to generate a time series pattern stability graph; Perform spatial conduction path analysis on the feature response sequence, construct a feature conduction directed graph based on the activation order of the hierarchical feature response subsequence, identify the feature blocking nodes and redundant conduction branches in the conduction path through graph structure analysis, and generate a spatial conduction efficiency evaluation matrix; Construct a multi-dimensional feature response quality evaluation model combining the time series pattern stability graph and the spatial conduction efficiency evaluation matrix, the multi-dimensional feature response quality evaluation model contains a time series 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; Output the feature response quality index of each dimension through the multi-dimensional feature response quality evaluation model, identify the time series fragile area with a time series consistency index lower than a baseline value, the conduction blocking area with a spatial conduction efficiency index lower than a baseline value, and the anti-interference weak area with a robustness index lower than a baseline value; For the time series fragile area, generate parameter feature time series alignment suggestions including feature response delay compensation mechanism and cross-window pattern correlation weight adjustment strategy based on pattern evolution trend analysis; for the conduction blocking area, generate structure correlation optimization direction including adjacent level connection strengthening of blocking nodes and node deletion scheme of redundant branches based on path analysis of the feature conduction directed graph; for the anti-interference weak area, generate feature enhancement adaptation suggestions including abnormal feature sensitive level activation function type replacement and feature mapping space expansion strategy based on the response decay characteristics of the abnormal feature subset; Generate optimization suggestion information containing multi-dimensional optimization priority sorting based on the time series alignment suggestions, the structure correlation optimization direction, and the feature enhancement adaptation suggestions.

9. 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 federal learning framework as claimed in any one of claims 1-8.

10. 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 1-8.

Images