Consumer goods data security protection method and system based on federal learning

Abstract

The invention relates to the field of federated learning data security protection, in particular to a federated learning-based consumer goods data security protection method and system, and the method comprises the steps: enabling consumer goods production state data and federated learning gradient data of a production enterprise to generate a polynomial order through a prediction model, and enabling the prediction model to be constructed based on an LSTM architecture; generating a dynamic polynomial and a production enterprise local end encryption share based on the polynomial order and federated learning gradient data; and the federated learning local end uploads the plurality of production enterprise local end encryption shares to a public service platform, aggregates and decrypts the production enterprise local end encryption shares, generates federated learning gradient data, and uses the federated learning gradient data for updating a federated learning model. According to the federal learning consumer goods data security method, the order of balancing security and efficiency generated by introducing the LSTM through the polynomial is realized, and the federal learning consumer goods data security method considering privacy protection, security policy and lightweight calculation is realized.

Description

Technical Field

[0001] This invention relates to the field of data security protection in federated learning, and more particularly to a method and system for protecting consumer product data security based on federated learning. Background Technology

[0002] In the consumer goods manufacturing industry, manufacturers need to optimize production decisions such as demand forecasting and quality monitoring through cross-enterprise collaboration. However, the equipment parameters and quality inspection records they use involve trade secrets, and direct sharing could lead to privacy risks. Federated learning, as a distributed machine learning paradigm, allows local machines to collaboratively train models without sharing raw data, exchanging only model gradient information. However, traditional federated learning faces the following security challenges in the consumer goods industry: First, attackers can reverse engineer gradient data using techniques such as model inversion attacks to deduce the original production state, leading to the leakage of trade secrets. Second, secret sharing using polynomials of fixed order is vulnerable to collusion attacks, such as malicious nodes colluding to crack encrypted shares, also resulting in the leakage of trade secrets. Third, secure aggregation schemes such as homomorphic encryption have high computational overhead and are difficult to adapt to production line scenarios with high real-time requirements.

[0003] Therefore, how to achieve a federated learning data security approach that balances strong privacy protection, dynamic security strategies, and lightweight computing to support secure cross-enterprise collaboration in the consumer goods industry is a technical problem that needs to be solved. Summary of the Invention

[0004] To address this, the present invention provides a method and system for protecting consumer product data security based on federated learning. Through a dynamic polynomial masking mechanism and lightweight Lagrange aggregation, it enables manufacturers to upload only the polynomial share after encryption masking. The public service platform decrypts and restores the gradient through aggregation, and the original gradient data is invisible throughout the process, completely blocking model inversion attacks. Furthermore, by introducing a polynomial to the order of LSTM generation that balances security and efficiency, it resists replay attacks, statistical analysis attacks, and collusion attacks. Thus, it achieves a federated learning method for protecting consumer product data that balances strong privacy protection, dynamic security strategies, and lightweight computation.

[0005] To achieve the above objectives, this invention proposes a consumer product data security protection method based on federated learning, comprising: Acquire consumer product production status data from manufacturing enterprises, and generate a multinomial order by combining the consumer product production status data and federated learning gradient data through a prediction model, wherein the prediction model is built based on an LSTM architecture. A dynamic polynomial is generated based on the polynomial order and the federated learning gradient data, and the local encryption share of the production enterprise is calculated based on the dynamic polynomial. Multiple local terminals of the federated learning process upload encrypted portions from multiple local terminals of the aforementioned production enterprises to the public service platform. The consumer goods public service platform aggregates and decrypts the encrypted portions from multiple local terminals of the aforementioned production enterprises to generate the federated learning gradient data, and uses the federated learning gradient data to update the federated learning model.

[0006] Furthermore, the process of generating polynomial orders from consumer goods production status data and federated learning gradient data through a prediction model includes: The consumer goods production status data and federated learning gradient data are fused to construct input features; The input features are used to generate output hidden states through the gating mechanism of the prediction model; The output hidden state is used to generate the polynomial order through linear combination and rounding.

[0007] Furthermore, the process of generating the output hidden state from the input features through the gating mechanism of the prediction model includes: The gating mechanism calculates the input gate, output gate, and candidate state vector based on the input features and the hidden state of the previous time step. The gating mechanism calculates the forget gate based on the input features, the hidden state of the previous time step, and the cell state update value of the previous time step; The gating mechanism calculates the current time step cell state update value based on the input gate, candidate state vector, forget gate, and the cell state update value of the previous time step. The gating mechanism calculates the hidden state at the current time step based on the cell state update value at the current time step and the output gate; The gating mechanism iteratively calculates the hidden state of the current time step until the iteration upper limit threshold is exceeded, and then uses the hidden state of the current time step as the output hidden state.

[0008] Furthermore, the gating mechanism's process of calculating the candidate state vector based on the input features and the hidden state of the previous time step includes: The concatenated vector of the input features and the hidden state of the previous time step is used to generate nonlinear correlation features through a convolution operation; The nonlinear correlation features are used to generate candidate state vectors through a multilayer perceptron.

[0009] Furthermore, the process of generating a polynomial order from the output hidden state through linear combination and rounding includes: The output hidden state is used to generate a prediction order ratio through a fully connected layer; The polynomial order is calculated based on the predicted order ratio, the upper limit of the polynomial order, and the lower limit of the polynomial order. The order of the polynomial is constrained based on the upper limit and lower limit of the polynomial order.

[0010] Furthermore, the process of obtaining consumer goods production status data from manufacturing enterprises includes: By acquiring resource usage data of industrial controllers in manufacturing enterprises through the Internet of Things, as well as physical status data collected by edge sensors; The consumer goods production status data is constructed based on the resource occupancy data and physical status data. The resource usage data includes the CPU utilization rate of the production equipment controller, memory usage, local federated learning data volume, federated learning communication latency, and federated learning stage encoding.

[0011] In the above scheme, the prediction model is built on LSTM. Its gating mechanism can effectively handle the time-series characteristics of consumer goods production status data, including the temporal fluctuations of industrial controller resource usage and sensor physical states. By capturing long-term dependencies in the data, LSTM can more accurately generate polynomial orders that match the actual production scenario, ensuring that the dynamic polynomial is highly consistent with the data features and constructing multi-dimensional production status data. Rich input features provide the LSTM model with more comprehensive scenario information, which helps to generate more reasonable polynomial orders, thereby optimizing the adaptability of dynamic encryption.

[0012] Furthermore, the process of generating a dynamic polynomial based on the polynomial order and federated learning gradient data includes: The public identity data of the local end of the federated learning is multiplied by random coefficients to construct the random term of the polynomial order; The dynamic polynomial is generated by summing the random terms and the federated learning gradient data.

[0013] Furthermore, the process of generating federated learning gradient data includes: The consumer goods public service platform aggregates and calculates the encrypted shares of multiple local terminals of the aforementioned manufacturers to generate aggregated shares; The base value is calculated using the Lagrange interpolation method for multiple public identity data from multiple local terminals of multiple production enterprises; The federated learning gradient data is calculated based on the aggregation share and the base value.

[0014] Furthermore, the process of generating aggregated shares includes: performing aggregate calculations on at least two encrypted shares of the current production enterprise local end distributed to other production enterprise local ends to generate the aggregated share of the current production enterprise local end; The process of calculating federated learning gradient data includes: summing up the aggregation shares of all the current production enterprise local terminals to determine the total aggregation share, and using the product of the total aggregation share and the base value as the federated learning gradient data.

[0015] The above scheme achieves efficient reconstruction of federated learning gradient data by using Lagrange interpolation, which has a significantly lower computational complexity than homomorphic encryption. It also enables decentralized collaboration among multiple manufacturing enterprises to securely share federated learning gradient data on their local terminals through a public service platform, thus promoting the progress of federated learning in the consumer goods manufacturing industry in terms of real-time performance and compliance.

[0016] The present invention also provides an intelligent guidance system based on multimodal perception, which applies the aforementioned intelligent guidance method based on multimodal perception.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By using a dynamic polynomial masking mechanism and lightweight Lagrange aggregation, the system enables manufacturers to upload only the polynomial share after encryption. The public service platform decrypts and restores the gradient through aggregation, and the original gradient data is invisible throughout the process, completely blocking model inversion attacks. Furthermore, by introducing a polynomial to balance security and efficiency in LSTM generation, the system can resist replay attacks, statistical analysis attacks, and collusion attacks. This results in a federated learning consumer product data security method that balances strong privacy protection, dynamic security strategies, and lightweight computation.

[0018] 2. The prediction model is built on LSTM, whose gating mechanism effectively handles the time-series characteristics of consumer goods production status data, including the temporal fluctuations in industrial controller resource usage and sensor physical states. By capturing long-term dependencies in the data, LSTM can more accurately generate polynomial orders that match the actual production scenario, ensuring a high degree of fit between the dynamic polynomial and data features, and constructing multi-dimensional production status data. Rich input features provide the LSTM model with more comprehensive scenario information, helping to generate more reasonable polynomial orders and thus optimizing the adaptability of dynamic encryption.

[0019] 3. By using Lagrange interpolation, which has a significantly lower computational complexity than homomorphic encryption, we have achieved efficient reconstruction of federated learning gradient data. This enables a decentralized collaborative approach where multiple manufacturing enterprises can securely share federated learning gradient data on their local end through a public service platform, thus promoting the progress of federated learning in the consumer goods manufacturing industry in terms of real-time performance and compliance. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the consumer product data security protection method based on federated learning, as described in an embodiment of the present invention. Figure 2This is a flowchart illustrating the process of generating the order of the polynomial in the federated learning-based consumer product data security protection method according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the gating mechanism of the consumer product data security protection method based on federated learning, as described in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the encryption and decryption process of the consumer product data security protection method based on federated learning, according to an embodiment of the present invention. Detailed Implementation

[0021] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0022] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0024] Furthermore, it should be noted that, in the description of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0025] like Figures 1 to 4 As shown, this invention provides an intelligent guidance method and system based on multimodal perception. The company's basic data, generated through user-input multimodal data conversion, and the regional feature data from the database are fused through a gating unit. This achieves the identification of complex relationships between regional feature data and company basic data, enabling the intelligent guidance files generated through semantic recognition to be adaptable to different regions. This simplifies the process of manually editing employee guidance files and improves the efficiency of the enterprise service platform in providing regionalized guidance to enterprise employees.

[0026] like Figures 1 to 4As shown, this embodiment proposes a consumer product data security protection method based on federated learning, including: Acquire consumer product production status data from manufacturing enterprises, and generate a multinomial order by combining the consumer product production status data and federated learning gradient data through a prediction model, wherein the prediction model is built based on an LSTM architecture. A dynamic polynomial is generated based on the polynomial order and the federated learning gradient data, and the local encryption share of the production enterprise is calculated based on the dynamic polynomial. Multiple local terminals of the federated learning process upload encrypted portions from multiple local terminals of the aforementioned production enterprises to the public service platform. The consumer goods public service platform aggregates and decrypts the encrypted portions from multiple local terminals of the aforementioned production enterprises to generate the federated learning gradient data, and uses the federated learning gradient data to update the federated learning model.

[0027] Understandably, by adopting a federated learning architecture where manufacturers generate encrypted shares locally and the public platform aggregates and decrypts them, manufacturers only upload share data that has been dynamically encrypted using a multinomial, rather than the original production status data. This avoids the risk of leakage of core production data (such as capacity and process parameters) from the source and meets the data privacy protection needs of multi-entity collaborative scenarios in the consumer goods industry.

[0028] Understandably, by using an LSTM prediction model to combine production status data and gradient data to generate a polynomial order, a dynamic polynomial encryption scheme can be constructed. This scheme can adjust encryption parameters according to temporal changes such as production capacity fluctuations and equipment status switching. Compared with traditional fixed polynomial encryption methods, it can more flexibly meet the security needs of complex production scenarios and reduce the possibility of the encryption scheme being reverse-engineered.

[0029] like Figure 2 As shown, the process of generating a polynomial order from consumer goods production status data and federated learning gradient data through a prediction model further includes: The consumer goods production status data and federated learning gradient data are fused to construct input features; The input features are used to generate output hidden states through the gating mechanism of the prediction model; The output hidden state is used to generate the polynomial order through linear combination and rounding.

[0030] Understandably, fusing consumer goods production status data, including equipment operating parameters and capacity fluctuations, with federated learning gradient data to construct input features achieves bidirectional information integration between the actual state of the production scenario and the model training requirements. This multi-source fusion mechanism can more accurately capture the correlation features between the production process and model updates, providing more comprehensive information support for subsequently generating polynomials adapted to the actual scenario, thereby enhancing the matching degree between encryption strategies and actual needs from the source.

[0031] like Figure 3 As shown, the process of generating the output hidden state from the input features through the gating mechanism of the prediction model further includes: The gating mechanism calculates the input gate, output gate, and candidate state vector based on the input features and the hidden state of the previous time step. The gating mechanism calculates the forget gate based on the input features, the hidden state of the previous time step, and the cell state update value of the previous time step; The gating mechanism calculates the current time step cell state update value based on the input gate, candidate state vector, forget gate, and the cell state update value of the previous time step. The gating mechanism calculates the hidden state at the current time step based on the cell state update value at the current time step and the output gate; The gating mechanism iteratively calculates the hidden state of the current time step until the iteration upper limit threshold is exceeded, and then uses the hidden state of the current time step as the output hidden state.

[0032] Understandably, the computation of the input gate, output gate, and candidate state vector deeply integrates the current input features with the hidden state of the previous time step, forming a two-way information interaction between the current input and historical memory. The input gate dynamically filters valuable features in the current input for generating the encryption order, such as abnormal equipment parameters and key gradient directions. The candidate state vector extracts latent features of the input through the tanh function. The output gate further controls the amount of information in the final output. The three work together to avoid redundant information interference, such as instantaneous equipment jitter data, while preserving the continuous features of time-series data, such as the trend of declining production capacity, providing more accurate intermediate state support for the subsequent generation of polynomial orders that fit the production scenario.

[0033] like Figure 2 As shown, further, the gating mechanism calculates the candidate state vector based on the input features and the hidden state of the previous time step, including: The concatenated vector of the input features and the hidden state of the previous time step is used to generate nonlinear correlation features through a convolution operation; The nonlinear correlation features are used to generate candidate state vectors through a multilayer perceptron.

[0034] It is understandable that complex nonlinear relationships often exist in the concatenated vector of input features and the hidden state (historical key information) from the previous time step. Processing the concatenated vector through convolution operations can effectively capture these local or global nonlinear feature patterns, such as the synergistic changes in production states and gradient data within a specific time window. Compared to the linear combination of traditional fully connected layers, this significantly improves the ability to represent deep data relationships, providing a richer feature foundation for subsequent generation of candidate state vectors. The nonlinear relationship features output by the convolution operation are further processed by a multilayer perceptron (MLP). Its multilayer neuron structure enables hierarchical abstraction of features. Bottom-level neurons focus on extracting basic features, such as the specific magnitude of device anomalies; middle-level neurons integrate cross-feature relationships, such as the correlation strength between anomalies and gradient directions; and top-level neurons abstract high-level features strongly correlated with the generation of encryption orders, such as the influence weight of anomaly patterns on the multinomial order.

[0035] like Figure 3 As shown, the process of generating a polynomial order from the output hidden state through linear combination and rounding includes: The output hidden state is used to generate a prediction order ratio through a fully connected layer; The polynomial order is calculated based on the predicted order ratio, the upper limit of the polynomial order, and the lower limit of the polynomial order. The order of the polynomial is constrained based on the upper limit and lower limit of the polynomial order.

[0036] Understandably, the introduction of upper and lower limits for the polynomial order helps avoid over-encryption and under-encryption. Specifically, the upper limit constrains the maximum possible order, preventing a surge in computational complexity due to excessively high order and reducing the computing power burden on the local end of the production enterprise. The lower limit ensures that the order is not lower than the basic security requirements, preventing insufficient encryption strength and vulnerability to reverse engineering due to excessively low order, thus protecting privacy and security during data transmission.

[0037] like Figure 2 As shown, the process of obtaining consumer goods production status data from manufacturing enterprises further includes: By acquiring resource usage data of industrial controllers in manufacturing enterprises through the Internet of Things, as well as physical status data collected by edge sensors; The consumer goods production status data is constructed based on the resource occupancy data and physical status data. The resource usage data includes the CPU utilization rate of the production equipment controller, memory usage, local federated learning data volume, federated learning communication latency, and federated learning stage encoding.

[0038] Specifically, the production equipment controller includes a PLC controller, a SCADA system, and an industrial computer, which communicate with the local end of the production enterprise via Modbus / TCP, MQTT, and EtherNet / IP, respectively, for Internet of Things (IoT) purposes. The physical status data includes vibration spectrum, temperature fluctuation, and power ripple, which communicate with the local end of the production enterprise via industrial bus, RS485, and CAN bus, respectively, for IoT purposes.

[0039] Specifically, the input features can be represented as: In the formula, This represents the input feature at the current time step t. Indicates CPU utilization. Indicates memory usage. This indicates the amount of locally federated learning data. This indicates a delay in federal learning communications. This indicates the production stage code, for example, fermentation=1, filling=2, packaging=3, Vib, degree, This represents physical state data, where Vib represents time-series data of vibration spectrum acquired by IEPE accelerometers after preprocessing and feature extraction, and degree represents time-series data of PT100 platinum resistance thermometers after preprocessing and feature extraction. This represents time-series data obtained from power supply ripple acquired by a high-voltage differential probe after preprocessing and feature extraction. This indicates the order of the polynomial.

[0040] Understandably, due to the strong coupling between the real-time nature of industrial control systems and production load, CPU utilization reflects the intensity of control logic execution, memory usage reflects the scale of production data, and production stage coding reflects the production status. For example, when a sensor malfunction causes the program to get stuck in a waiting loop, the extended business execution time leads to a surge in CPU utilization, and exceeding historical data cache limits causes a spike in memory usage. Therefore, CPU utilization and memory usage can ensure that the polynomial order of the dynamic polynomial adapts to the resource usage of the production equipment controller on the local end of the production enterprise. When production consumes a large amount of resources, a lower polynomial order is set to reduce the load on the controller.

[0041] Understandably, the local federated learning data volume and federated learning communication latency can achieve a balance between the security, efficiency and model quality of federated learning. When the federated learning communication latency is large, the polynomial order is automatically reduced to prioritize timeliness. When the data volume increases by a large factor, the polynomial order is reduced to ensure that the cost of a single training round does not exceed the limit.

[0042] Specifically, the gating mechanism can be represented as: In the formula, This represents the input feature at the current time step t. This indicates the hidden state at the previous time step t-1. The concatenated vector of the input features at the current time step t. This indicates the hidden state at the previous time step t-1. The concatenated vector of the input features at the current time step t and the concatenated vector of the cell state update values ​​from the previous time step. This represents the forget gate at the current time step t. This represents the input gate for the current time step t. This represents the output gate at the current time step t. This represents the candidate state vector at the current time step t. This represents the cell state update value at the current time step. This represents the cell state update value from the previous time step. When the iteration proceeds to the next time step, the cell state update value at the current time step is automatically updated to the cell state update value from the previous time step. Indicates the order of the polynomial. This represents the sigmoid function. Represents the tanh function. This indicates element-wise multiplication; `round` represents the floor function. , , , , , These represent the weight matrices for the corresponding gates. , , , , , These represent the bias vectors for the corresponding gates. The forget gate... By introducing peephole connections, direct access to the cell state from the previous time step is achieved, enabling gating to more accurately determine what content needs to be retained or forgotten. The standard LSTM cell state update is as follows: This embodiment, based on this, adds a nonlinear transformation through convolution operation. By using a linear multilayer perceptron, the ability to capture complex situations is improved. Among these... This indicates that the fully connected layer outputs the prediction order proportionally through the sigmoid function, with the following constraint on the polynomial order: Where 2 is the preferred lower limit of the polynomial order, 6 is the preferred upper limit of the polynomial order, and N is a set value ranging from 0.1 to 1.25.

[0043] For example: if the cell state at the previous time step is 0.8, the forget gate output is 0.4, the input gate output is 0.7, and the candidate state vector is 0.6, then... [0.32 (retaining some temperature stability memory), 0.08 (weakening vibration memory), 0.04 (almost forgetting ripple memory)], in the filling stage, since temperature is no longer a core indicator, and equipment switching causes load changes, only 32% of the fermentation temperature stability memory is retained, while the fermentation vibration and ripple memory are greatly weakened because they are unrelated to the filling stage. =0.7×[0.6,0.5,0.3]=[0.42 (high CPU load feature), 0.35 (high memory usage feature), 0.21 (network latency feature)]. To achieve the goal of focusing on equipment load (CPU / memory) and network status during the filling stage, these three types of features in the candidate state vector are amplified by the input gate (70% weight) and become the core of the new cell state.

[0044] Preferably, the model parameters of the prediction model include: 9 input layer nodes, 0.2 Dropout layer ratio, 16 fully connected layer nodes, AdamW optimizer, 0.01 learning rate, 32 batch size, Huber loss function, and 200 training cycles (iteration upper limit threshold).

[0045] In the above scheme, the prediction model is built on LSTM. Its gating mechanism can effectively handle the time-series characteristics of consumer goods production status data, including the temporal fluctuations of industrial controller resource usage and sensor physical states. By capturing long-term dependencies in the data, LSTM can more accurately generate polynomial orders that match the actual production scenario, ensuring that the dynamic polynomial is highly consistent with the data features and constructing multi-dimensional production status data. Rich input features provide the LSTM model with more comprehensive scenario information, which helps to generate more reasonable polynomial orders, thereby optimizing the adaptability of dynamic encryption.

[0046] like Figure 4 As shown, the process of generating a dynamic polynomial based on the polynomial order and federated learning gradient data further includes: The public identity data of the local end of the federated learning is multiplied by random coefficients to construct the random term of the polynomial order; The dynamic polynomial is generated by summing the random terms and the federated learning gradient data.

[0047] Specifically, the dynamic polynomial can be expressed as: In the formula, Let represent the dynamic polynomial belonging to the k-th gradient component of the local terminal i of the current production enterprise. This represents the value of the k-th gradient component belonging to the local terminal i of the current production enterprise. This indicates the first to the last gradient component belonging to the k-th gradient component. -1 random coefficient, The first to the second power of the public identity data belonging to the local terminal i of the current production enterprise -1 power, This represents the remainder operation, i.e. Divide by a large prime number Taking the remainder makes it difficult for attackers to decompose the data. Alternatively, one can compute the inverse to solve the dynamic polynomial. Among these, Indicates the order of the polynomial. The random item is represented by the public identity data, which is preferably the public identity data ID of the current production enterprise local terminal i. The public identity data ID is a fixed integer known to other local terminals. The federated learning gradient data contains multiple gradient components. The random coefficient is a coefficient randomly selected from a set finite field to ensure the randomness of encryption.

[0048] Specifically, to protect the federated learning gradient data of the current production enterprise's local terminal i from being leaked by the consumer goods public service platform's server or other clients, a dynamic polynomial is calculated for each gradient component of the federated learning towards other production enterprise's local terminals j. This dynamic polynomial is then used as the encrypted share of the production enterprise's local terminal. ,in This represents the encrypted share of the local terminal of the current production enterprise i, belonging to the k-th gradient component, distributed to the local terminals of other production enterprises j. This represents the public identity data belonging to the k-th gradient component, which includes the local j-th endpoint of other production enterprises. The dynamic polynomial calculated locally at the current production enterprise (i) is substituted into the equation. Preferably, the encrypted portion at the production enterprise's local end is transmitted using the Paillier or Shamir encryption algorithm to ensure transmission security.

[0049] like Figure 4 As shown, the process of generating federated learning gradient data further includes: The consumer goods public service platform aggregates and calculates the encrypted shares of multiple local terminals of the aforementioned manufacturers to generate aggregated shares; The base value is calculated using the Lagrange interpolation method for multiple public identity data from multiple local terminals of multiple production enterprises; The federated learning gradient data is calculated based on the aggregation share and the base value.

[0050] Specifically, the cryptographic share is a trust anchor for secure multi-party computation in threshold cryptography-based federated learning, and the dynamic polynomial is generated through the SM9 curve.

[0051] Furthermore, the process of generating aggregated shares includes: performing aggregate calculations on at least two encrypted shares of the current production enterprise local end distributed to other production enterprise local ends to generate the aggregated share of the current production enterprise local end; The process of calculating federated learning gradient data includes: summing up the aggregation shares of all the current production enterprise local terminals to determine the total aggregation share, and using the product of the total aggregation share and the base value as the federated learning gradient data.

[0052] Specifically, using The properties of the base values ​​are calculated using the Lagrange interpolation method. The process can be represented as: Define the Lagrange polynomial: ,satisfy Calculate the value of the basis polynomial at x=0: In the above formula, This represents the total number of other production enterprise local terminals j that received the encrypted share from the production enterprise's local terminal. Indicates the base value. This represents the sum of the dynamic polynomials of the k-th gradient components distributed from the local terminals i of the current production enterprise to the local terminals j of other production enterprises, which is the sum of all global dynamic polynomials. This represents the remainder operation, i.e. Divide by a large prime number Take the remainder. This represents the sum of the values ​​of the k-th gradient components of the n local terminals i of the current production enterprises, that is, the constant term of the sum of all global dynamic polynomials, which is the federated learning gradient data. These represent the interpolated public identity data x and the public identity data of other production enterprises' local terminals j, respectively. And the public identity data of the local end of the m-th production enterprise The corresponding Lagrange basis function value. Therefore, the base value can be determined by substituting the public identity data, preferably the public identity data ID, into the Lagrange basis function value to recover the value of the constant term.

[0053] Specifically, the process of calculating federated learning gradient data can be represented as: In the formula, This represents the sum of the values ​​of the k-th gradient components of the n local terminals i of the current production enterprises, that is, the constant term of the sum of all global dynamic polynomials, which is the federated learning gradient data. This represents the total number of other production enterprise local terminals j that received the encrypted share from the production enterprise's local terminal. This represents the aggregate share belonging to the k-th gradient component, sent from the local terminal j of other manufacturing enterprises to the server of the consumer goods public service platform. This represents the remainder operation, i.e. Divide by a large prime number The remainder is taken, where the aggregated share is the sum of the encrypted shares received by the local end of all production enterprises from the server of the consumer goods public service platform, i.e. ,in This represents the encrypted share of the local end of the current production enterprise i, which belongs to the k-th gradient component, distributed to the local end of other production enterprises j.

[0054] Therefore, the process described by the above formula involves calculating the aggregation share. Calculate the numerical sum of the gradient components. Furthermore, through Lagrange interpolation... Points The value of the reconstructed dynamic polynomial is decrypted, and the federated learning gradient data is decrypted, thereby preventing the secret sharing in polynomials of fixed order from being easily cracked by malicious nodes in a joint effort.

[0055] The above scheme reduces decryption computational overhead by using lightweight polynomial operations and Lagrange interpolation decryption, and achieves this through the aggregation of multiple participants on the local end. This embodiment also provides a consumer product data security protection system based on federated learning, wherein the consumer product data security protection system applies the aforementioned consumer product data security protection method based on federated learning.

[0056] In this embodiment, a dynamic polynomial masking mechanism and lightweight Lagrange aggregation are used to enable manufacturers to upload only the polynomial share after encryption. The public service platform decrypts and restores the gradient through aggregation, keeping the original gradient data invisible throughout the process, thus completely blocking model inversion attacks. Furthermore, the introduction of a polynomial to the LSTM-generated order balances security and efficiency, resisting replay attacks, statistical analysis attacks, and collusion attacks. This achieves a federated learning method for consumer product data security that balances strong privacy protection, dynamic security strategies, and lightweight computation. The prediction model is built on LSTM, whose gating mechanism effectively handles the time-series characteristics of consumer product production status data, including the temporal fluctuations in industrial controller resource usage and sensor physical states. By capturing long-term dependencies in the data, LSTM can more accurately generate polynomial orders that match the actual production scenario, ensuring a high degree of fit between the dynamic polynomial and data features, and constructing multi-dimensional production status data. Rich input features provide the LSTM model with more comprehensive scenario information, helping to generate more reasonable polynomial orders and thus optimizing the adaptability of dynamic encryption. By using Lagrange interpolation, which has a computational complexity significantly lower than homomorphic encryption, efficient reconstruction of federated learning gradient data is achieved. This enables a decentralized collaborative approach where multiple manufacturing enterprises can securely share federated learning gradient data on their local end through a public service platform, thus promoting the progress of federated learning in the consumer goods manufacturing industry in terms of real-time performance and compliance.

[0057] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0058] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

[0059] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for protecting consumer product data security based on federated learning, characterized in that, include: Acquire consumer product production status data from manufacturing enterprises, and generate a multinomial order by combining the consumer product production status data and federated learning gradient data through a prediction model, wherein the prediction model is built based on an LSTM architecture. A dynamic polynomial is generated based on the polynomial order and the federated learning gradient data, and the local encryption share of the production enterprise is calculated based on the dynamic polynomial. Multiple local terminals of the federated learning process upload encrypted portions from multiple local terminals of the aforementioned production enterprises to the public service platform. The consumer goods public service platform aggregates and decrypts the encrypted portions from multiple local terminals of the aforementioned production enterprises to generate the federated learning gradient data, and uses the federated learning gradient data to update the federated learning model.

2. The consumer product data security protection method based on federated learning according to claim 1, characterized in that, The process of generating polynomial orders from consumer goods production status data and federated learning gradient data through a prediction model includes: The consumer goods production status data and federated learning gradient data are fused to construct input features; The input features are used to generate output hidden states through the gating mechanism of the prediction model; The output hidden state is used to generate the polynomial order through linear combination and rounding.

3. The consumer product data security protection method based on federated learning according to claim 2, characterized in that, The process of generating output hidden states from input features through the gating mechanism of a prediction model includes: The gating mechanism calculates the input gate, output gate, and candidate state vector based on the input features and the hidden state of the previous time step. The gating mechanism calculates the forget gate based on the input features, the hidden state of the previous time step, and the cell state update value of the previous time step; The gating mechanism calculates the current time step cell state update value based on the input gate, candidate state vector, forget gate, and the cell state update value of the previous time step. The gating mechanism calculates the hidden state at the current time step based on the cell state update value at the current time step and the output gate; The gating mechanism iteratively calculates the hidden state of the current time step until the iteration upper limit threshold is exceeded, and then uses the hidden state of the current time step as the output hidden state.

4. The consumer product data security protection method based on federated learning according to claim 3, characterized in that, The gating mechanism calculates the candidate state vector based on the input features and the hidden state of the previous time step, including the following process: The concatenated vector of the input features and the hidden state of the previous time step is used to generate nonlinear correlation features through a convolution operation; The nonlinear correlation features are used to generate candidate state vectors through a multilayer perceptron.

5. The consumer product data security protection method based on federated learning according to claim 2, characterized in that, The process of generating a polynomial order from the output hidden state through linear combination and rounding includes: The output hidden state is used to generate a prediction order ratio through a fully connected layer; The polynomial order is calculated based on the predicted order ratio, the upper limit of the polynomial order, and the lower limit of the polynomial order. The order of the polynomial is constrained based on the upper limit and lower limit of the polynomial order.

6. The consumer product data security protection method based on federated learning according to claim 1, characterized in that, The process of obtaining consumer goods production status data from manufacturing enterprises includes: By acquiring resource usage data of industrial controllers in manufacturing enterprises through the Internet of Things, as well as physical status data collected by edge sensors; The consumer goods production status data is constructed based on the resource occupancy data and physical status data. The resource usage data includes the CPU utilization rate of the production equipment controller, memory usage, local federated learning data volume, federated learning communication latency, and federated learning stage encoding.

7. The consumer product data security protection method based on federated learning according to claim 1, characterized in that, The process of generating a dynamic polynomial based on the polynomial order and federated learning gradient data includes: The public identity data of the local end of the federated learning is multiplied by random coefficients to construct the random term of the polynomial order; The dynamic polynomial is generated by summing the random terms and the federated learning gradient data.

8. The consumer product data security protection method based on federated learning according to claim 7, characterized in that, The process of generating federated learning gradient data includes: The consumer goods public service platform aggregates and calculates the encrypted shares of multiple local terminals of the aforementioned manufacturers to generate aggregated shares; The base value is calculated using the Lagrange interpolation method for multiple public identity data from multiple local terminals of multiple production enterprises; The federated learning gradient data is calculated based on the aggregation share and the base value.

9. The consumer product data security protection method based on federated learning according to claim 8, characterized in that, The process of generating aggregated shares includes: performing aggregate calculations on at least two encrypted shares of the current production enterprise local end that are distributed to other production enterprise local ends to generate aggregated shares of the current production enterprise local end; The process of calculating federated learning gradient data includes: summing up the aggregation shares of all the current production enterprise local terminals to determine the total aggregation share, and using the product of the total aggregation share and the base value as the federated learning gradient data.

10. A consumer product data security protection system based on federated learning, characterized in that, The consumer product data security protection system applies the consumer product data security protection method based on federated learning as described in any one of claims 1 to 9.

Images