Supply chain data security sharing method based on federated learning

Abstract

The application relates to the field of artificial intelligence and federal learning, and discloses a supply chain data security sharing method based on federal learning. The method constructs a decentralized multi-participant collaborative architecture, applies local differential privacy disturbance to local gradients in each iteration, and adopts a double mechanism of threshold secret sharing and Paillier homomorphic encryption to split, encrypt and transmit gradient shares; each participant reconstructs the global gradient through linear superposition and completes model updating locally. The application does not require a trusted third party, effectively prevents leakage of original data and gradient information, balances model utility and strong privacy protection, and is suitable for cross-enterprise supply chain intelligent collaboration scenarios.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence and federated learning, and particularly relates to a supply chain data security sharing method based on federated learning. BACKGROUND

[0002] With the rapid development of digital economy, the demand for cross-enterprise data sharing in supply chain collaborative management is increasingly urgent. The traditional centralized data integration method requires each participant to upload the original business data to the central platform for joint modeling and analysis, which is easy to lead to the leakage of sensitive business information and is difficult to meet the compliance requirements of the Data Security Law and the Personal Information Protection Law. In order to realize collaborative intelligence on the premise of protecting data privacy, federated learning as a distributed machine learning paradigm is introduced into the supply chain scenario, and its core idea is to keep the original data locally and only exchange model parameters or gradient updates, thereby avoiding the direct transmission of sensitive information.

[0003] The supply chain data security sharing method based on federated learning aims to train a global prediction model through multi-party collaboration without exposing private data. This method usually uses a central server to coordinate the local model training of each participant and aggregates the uploaded gradients to update the global model, and then distributes the updated model to each party for the next round of iteration.

[0004] In the prior art, although federated learning avoids the direct transmission of original data, the gradients uploaded by participants may still imply the statistical characteristics or even specific contents of the original data. Research shows that attackers can reconstruct part of the original input data from shared gradients through gradient inversion, member inference or model stealing. Especially in the supply chain scenario, since business data is highly structured, sparse and semantically related, the information carried by the gradient is more likely to be reverse analyzed.

[0005] More seriously, as the coordinator, the central server itself may have malicious motives or be subject to intrusion, and it can use the received gradient information to implement targeted reasoning attacks, thereby inferring the business secrets of a party. In addition, relying solely on differential privacy or homomorphic encryption and other single privacy enhancement technologies often cannot balance between security and model utility: the noise introduced by the former will significantly reduce the model accuracy, and the latter will bring unacceptable computational and communication overhead, making it difficult to support the performance requirements of real-time collaboration in the supply chain. Therefore, there is an urgent need for a security sharing mechanism that can effectively block the gradient information leakage path and also guarantee the efficiency and accuracy of model training. SUMMARY

[0006] The application provides a supply chain data security sharing method based on federated learning, aiming to solve the privacy leakage risk caused by the central server stealing gradient information to infer the original data in federated learning. The method builds a decentralized multi-participant collaborative architecture, introduces a local differential privacy perturbation mechanism and a homomorphic encryption double protection strategy, and designs a gradient aggregation protocol based on secret sharing, ensuring that the model training and updating are completed without exposing any original data and intermediate gradients, thereby realizing the safe co-construction and efficient utilization of the joint intelligent model by each participant in the supply chain under the condition of data isolation.

[0007] The application provides a supply chain data security sharing method based on federated learning, comprising:

[0008] Deploying a local machine learning model with the same structure on each supply chain participant side, and setting unified model structure parameters;

[0009] In each round of federated learning iteration, each participant performs forward propagation and backward propagation calculation based on its local supply chain business data to generate a local gradient vector;

[0010] Applying random noise perturbation satisfying the local differential privacy constraint to the local gradient vector to obtain a perturbed gradient vector;

[0011] The perturbed gradient vector is divided into multiple shares according to the threshold secret sharing scheme, and each share is sent to other participants using the homomorphic encryption mechanism;

[0012] After each participant receives the encrypted gradient shares from all other participants, it performs linear superposition on all received shares according to the preset aggregation rule to reconstruct the encrypted aggregation result of the global gradient;

[0013] Based on the encrypted aggregation result of the global gradient, the model parameter update operation is performed locally to generate the local model required for the next iteration;

[0014] Repeat the above steps until the model converges or the maximum number of iterations is reached, and finally each participant obtains a consistent global model copy for supply chain prediction, inventory optimization or demand analysis tasks.

[0015] Preferably, the random noise perturbation satisfying the local differential privacy constraint is applied to the local gradient vector to obtain a perturbed gradient vector, comprising:

[0016] According to the preset privacy budget , determine the Laplace noise scale parameter :

[0017] ;

[0018] a gradient sensitivity, which is an upper bound of the L1 norm of the difference between the gradient vectors generated by any two adjacent data sets;

[0019] adding independent random noise following Laplace distribution to each dimension of the local gradient vector to form a perturbed gradient vector.

[0020] Preferably, the perturbed gradient vector is divided into multiple shares according to a threshold secret sharing scheme, including:

[0021] setting the total number of participants as and the reconstruction threshold as , wherein is less than and greater than half of

[0022] For each element of the perturbed gradient vector, a random polynomial with a degree of is constructed, whose constant term is the value of the element, and the remaining coefficients are uniformly randomly selected from a finite field;

[0023] The values of the polynomial at non-zero integer points are taken as shares and distributed to other participants respectively.

[0024] Preferably, each share is sent to other participants using a homomorphic encryption mechanism, including:

[0025] Each participant generates a pair of Paillier public and private keys in advance and broadcasts the public key to all other participants;

[0026] The sender encrypts the gradient share using the public key of the receiver before transmission;

[0027] The receiver decrypts the ciphertext share using its own private key to obtain the plaintext share, which is used for subsequent aggregation calculation.

[0028] Preferably, after each participant receives encrypted gradient shares from all other participants, it performs linear superposition on all received shares according to a predetermined aggregation rule to reconstruct the encrypted aggregation result of the global gradient, including:

[0029] Each participant performs element-wise homomorphic summation on all encrypted gradient shares it holds to obtain partial aggregation ciphertext;

[0030] When at least participants complete share exchange, any participants can recover the complete global perturbed gradient vector by Lagrange interpolation formula, which is the arithmetic mean of the local perturbed gradient of all participants.

[0031] ​​Preferably, the encryption aggregation result based on the global gradient performs a local model parameter update operation to generate a local model required for the next round of iteration, comprising:

[0032] Each participant locally maintains a momentum buffer, fuses the reconstructed global disturbance gradient with a historical momentum term by a preset decay factor, and generates an update direction;

[0033] Combined with a preset learning rate, the current local model parameters are synchronously adjusted.

[0034] Preferably, the supply chain business data includes purchase order records, logistics transportation status, warehouse inventory level, sales terminal flow and supplier credit rating information; the local machine learning model is a deep neural network, and the input layer dimension is matched with the feature dimension of the supply chain business data, and the output layer is configured as a classification head or a regression head according to the specific task type.

[0035] Preferably, before each round of iteration starts, each participant verifies and registers the identity of the participants in the current round through a blockchain consensus mechanism; only the participants who have been registered and authenticated can join the gradient sharing and aggregation process of the current round; the blockchain network records the participant list, aggregation timestamp and model version hash value of each round of iteration.

[0036] Preferably, if a participant has a communication interruption or response timeout during the gradient share transmission process, the remaining participants determine it as a failed node according to a preset fault tolerance strategy; in the subsequent aggregation stage, the contribution of the node is automatically excluded, and the reconstruction threshold of secret sharing is dynamically adjusted to maintain the system robustness; when the number of valid participants is less than , the current round is terminated and an abnormal alarm is triggered.

[0037] Preferably, the privacy budget is set to a larger value at the beginning of training to ensure the convergence speed of the model, and is gradually reduced at the later stage of training to enhance the privacy protection strength; the dynamic adjustment strategy is written into the smart contract by all participants through consensus in advance, and is automatically executed in each round of iteration.

[0038] Compared with the prior art, the present application has the following advantages:

[0039] 1. The present application completely abandons the traditional federated learning architecture which depends on a trusted central server, and eliminates the risk of single-point privacy leakage by constructing a completely decentralized multi-party cooperation mechanism.

[0040] 2. The introduction of the dual protection system of local differential privacy and homomorphic encryption strictly limits the possibility of any single participant or external attacker obtaining the original supply chain sensitive information through gradient inversion while ensuring the effectiveness of the model.

[0041] 3. The gradient aggregation protocol based on threshold secret sharing makes the reconstruction of the global gradient dependent on the cooperative efforts of the majority of participants, effectively resisting tampering or eavesdropping by malicious nodes.

[0042] 4. The combination of blockchain technology realizes the whole-process auditability and strong identity authentication, enhancing the transparency and anti-repudiation ability of the system.

[0043] 5. The dynamic privacy budget control strategy takes into account the training efficiency and long-term privacy security, and is suitable for complex and variable supply chain collaboration scenarios.

[0044] 6. The overall scheme realizes the unification of high security, strong robustness and good scalability without data centralization and sacrificing model accuracy, providing reliable technical support for cross-enterprise supply chain intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the overall technical scheme architecture diagram of the supply chain data security sharing method based on federated learning proposed by the application;

[0046] Figure 2 is the core principle framework diagram of the decentralized gradient aggregation and double privacy protection mechanism in the application;

[0047] Figure 3 is the logical flow framework diagram of local differential privacy disturbance and secret sharing share generation in the application;

[0048] Figure 4 is the logical flow framework diagram of gradient secure transmission and aggregation based on threshold secret sharing and homomorphic encryption in the application;

[0049] Figure 5 is the interaction relationship diagram of the blockchain-enabled participant identity authentication and training process auditability in the application;

[0050] Figure 6 is the dynamic privacy budget control and model iteration update collaborative control logical framework diagram in the application. DETAILED DESCRIPTION

[0051] Please refer to Figures 1 to 6The application provides a supply chain data security sharing method based on federated learning, which is characterized by constructing a fully decentralized multi-participant collaborative training architecture. Through the triple mechanism of local differential privacy disturbance, homomorphic encryption communication and threshold secret sharing aggregation, the enterprises in the supply chain can safely co-build a high-precision joint intelligent model without exchanging raw business data. The specific implementation steps of the method will be described in detail below. The method strictly follows the step sequence defined in the summary, and focuses on technical implementation details, data flow, parameter settings and exception handling mechanisms.

[0052] In the initialization phase, multiple supply chain participants each deploy a local machine learning model and set uniform model structure parameters. The supply chain participants include but are not limited to raw material suppliers, manufacturers, logistics carriers, warehousing service providers and retail terminal enterprises. Each participant loads the initial model weight on the local computing device. The model structure is determined by all participants through consultation before training starts, ensuring that the input layer dimension strictly matches the local supply chain business data feature dimension, and the output layer is configured as a classification head or a regression head according to the task type.

[0053] The supply chain business data includes purchase order records, logistics transportation status, warehouse inventory level, sales terminal flow and supplier credit rating information. All data is kept in local storage media and is not transmitted to any external entity. Model structure parameters include network layer number, number of neurons per layer, activation function type and loss function form. Once determined, they are fixed in the local execution environment of each participant until the training is completed.

[0054] After entering the first round of federated learning iteration, each participant performs forward propagation and backward propagation calculation based on its local supply chain business data to generate a local gradient vector.

[0055] Specifically, each participant divides its local business data into several batches, which are input into the local model in turn for forward reasoning, calculating the loss value between the predicted output and the true label. Then, the partial derivative of the loss function with respect to the model parameters is calculated layer by layer through the backpropagation algorithm, and finally converged into a local gradient vector with the same dimension as the model parameters. This gradient vector completely reflects the contribution of the current local data to the model update direction, but directly exposing the vector may lead to the reconstruction of the original data by the central server or other malicious participants through gradient inversion attack, so strict privacy protection measures must be taken.

[0056] A random noise disturbance satisfying the local differential privacy constraint is applied to the local gradient vector to obtain a disturbed gradient vector. The disturbance process is based on the pre-set privacy budget The execution firstly computes the gradient sensitivity, i.e. the upper bound of the L1 norm of the difference between the gradient vectors generated by any two adjacent data sets. The adjacent data sets are defined as the data sets that only differ by one record, and the gradient sensitivity is determined as the maximum influence amplitude of the model output on a single record. The Laplace noise scale parameter .

[0057] Subsequently, random noise following the Laplace distribution is independently added to each dimension of the local gradient vector. This operation ensures that the following conditions are met for any two adjacent data sets and , and any possible output , where denotes the perturbation mechanism. 、 denote the probabilities of the perturbation mechanism acting on the data set to obtain the output , and the perturbation mechanism acting on the adjacent data set to obtain the output . The perturbed gradient vector effectively masks the sensitive information of individual data while retaining statistical utility.

[0058] The perturbed gradient vector is divided into multiple shares, and each share is sent to other participants using a threshold secret sharing scheme. The threshold secret sharing scheme uses the Shamir polynomial interpolation method, sets the total number of participants to , and the reconstruction threshold to , where is less than and greater than half of , and a typical value is . .

[0059] For each element in the perturbed gradient vector, the sender constructs a random polynomial with degree , as the independent variable, and the remaining coefficients to are uniformly randomly selected from a sufficiently large finite field , is a large prime number with a bit length of no less than 2048 bits to ensure computational security. Subsequently, the sender calculates the value of the polynomial at (excluding its own identifier), obtaining shares , and sends each share assigned to the participating party. a random polynomial for the participating party; this process is performed in parallel for all dimensions of the gradient vector, forming a complete matrix of shares.

[0060] The transmission process of gradient shares employs a public-key encryption mechanism based on the Paillier homomorphic encryption system. Each participating party generates a pair of Paillier public and private keys in advance during system initialization, and the public key is broadcast to all other participating parties through a secure channel, while the private key is strictly kept secret in the local secure storage area. When a participating party needs to send a share to the participating party, the public key of the participating party is used to encrypt the share, generating a ciphertext . The operation of Paillier encryption is performed with the public key of the participating party.

[0061] The Paillier encryption system has additive homomorphism, that is , where is the public key modulus, , is the plaintext , and the ciphertext after Paillier encryption. This property provides the basis for subsequent aggregation calculations. The encrypted ciphertext is transmitted to the target participating party through standard network protocols, and the TLS1.3 protocol is used to ensure channel confidentiality and integrity during transmission.

[0062] After each participating party receives gradient shares from all other participating parties, it performs linear superposition on all received shares according to the pre-set aggregation rules to reconstruct the encrypted aggregation result of the global gradient. Specifically, each participating party maintains a receive buffer to temporarily store all encrypted shares received in this round. After all expected shares are received or the timeout determination is completed, the participating party performs element-wise homomorphic summation on all encrypted shares it holds.

[0063] Due to the additive homomorphism of Paillier, this summation operation can be completed in ciphertext state, obtaining a partial aggregation ciphertext . is the gradient share of the participating party to be sent to the participating party. This partial aggregation result represents the sum of all other participating parties' contributions to their shares, excluding itself. It is worth noting that due to the linear nature of secret sharing, the global perturbed gradient can be obtained by any ​The partial aggregation results of each participant are reconstructed. For the i-th participant. For the i-th participant.

[0064] The reconstruction process of the encrypted aggregation result of the global gradient is as follows: when at least participants survive and complete share exchange, any participants can cooperate to restore the complete global perturbation gradient vector. Each participant regards its decrypted partial aggregation plaintext value as a point , where is its unique identifier. For the i-th participant. The i-th element of the global perturbation gradient can be calculated using the Lagrange interpolation formula, and the i-th element of the global perturbation gradient can be calculated as:

[0065] ;

[0066] where is the index set of any surviving participants. The formula is executed in a finite field to ensure numerical stability. Since each , and , according to the reconstruction principle of the Shamir scheme, the above interpolation result is exactly , and divided by , i.e. the average gradient. This process does not require any central coordination node and is completely completed by the participants in point-to-point cooperation. For the i-th participant. The i-th element of the global perturbation gradient can be calculated using the Lagrange interpolation formula, and the i-th element of the global perturbation gradient can be calculated as: For the i-th participant. For the i-th participant. For the i-th participant. For the i-th participant. For the i-th participant. For the i-th participant. For the i-th participant. For the i-th participant. For the i-th participant.

[0067] Based on the encrypted aggregation result of the global gradient, a local model parameter update operation is performed locally to generate a local model required for the next round of iteration. The model parameter update operation uses a momentum gradient descent algorithm. Each participant locally maintains a momentum buffer with an initial value of a zero vector. In each round of iteration, the reconstructed global perturbation gradient​ with the historical momentum term by a preset decay factor weighted fusion, generating an update direction Subsequently, the momentum buffer is updated as assigning the value of to the momentum buffer Finally, the local model parameters are updated as assigning the value of to , with

[0068] being a preset learning rate. This update procedure is executed synchronously at all participants' local, ensuring that the model parameters at each end remain consistent. The above steps are repeated until the model converges or reaches a preset maximum number of iterations, and finally each participant obtains a consistent copy of the global model for subsequent supply chain prediction, inventory optimization or demand analysis tasks. The convergence criterion uses the validation set loss stability criterion: each participant keeps a small part of the validation data that does not participate in training locally, and calculates the validation loss after each update; when the change amplitude of the validation loss is less than a preset threshold for K consecutive rounds, the model is determined to have converged. The maximum number of iterations

[0069] serves as a hard termination condition to prevent infinite loops. After training is complete, each participant can independently use their local model copy to perform inference tasks without further communication.

[0070] Before each iteration begins, each participant verifies and registers the identity of the participants in this round through a blockchain consensus mechanism. The blockchain is a permissioned consortium chain maintained by the core enterprises of the supply chain. Participants must hold a certified digital certificate to join the network. At the start of each training round, each potential participant submits a registration request to the blockchain, including their public key, IP address, and model version hash. The on-chain smart contract performs identity verification, and only authorized entities can be included in the list of participants in this round. This list, the aggregated timestamp, and the initial model hash value of this round are written into a new block, ensuring that the entire process is auditable. If a participant experiences a communication interruption or response timeout during the gradient share transmission process, the remaining participants will determine it to be a failed node according to a preset fault tolerance strategy. The timeout threshold τ is dynamically set according to network delay statistics, with a typical value of 5 seconds. The missing share of the failed node will prevent it from contributing to the aggregation, but as long as the number of surviving nodes is ≥

[0071] Privacy budget The initial value is set to a larger value to ensure faster model convergence, and then gradually decreased in the later stages of training to enhance privacy protection. This dynamic adjustment strategy is pre-written into the smart contract through consensus among all participants. For example, initially... Multiply by the decay factor after every 10 iterations. until Reduced to 0.1. At the start of each round, the smart contract automatically broadcasts the current... Based on this value, each participant recalculates the noise scale parameters. This strategy allows for larger gradient perturbations in the early stages of training to accelerate convergence, while tightening privacy constraints in the later stages to improve long-term security, achieving a dynamic balance between utility and privacy.

[0072] In summary, this method completely eliminates the single point of failure risk of a central server through a decentralized architecture, and constructs a defense-in-depth system by combining local differential privacy, homomorphic encryption, and threshold secret sharing. All data is always kept locally, intermediate gradients are transmitted in encrypted shares after double perturbation and partitioning, and global model updates rely on the collaboration of most participants. Malicious nodes cannot reconstruct valid information. The blockchain mechanism strengthens identity authentication and operation traceability, and dynamic privacy control adapts to the entire training process. The overall solution protects supply chain data sovereignty while achieving highly secure and robust intelligent collaboration.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A supply chain data security sharing method based on federated learning, characterized in that, The method comprises the following steps: Deploying local machine learning models with the same structure on each supply chain participant side, and setting unified model structure parameters; In each round of federated learning iteration, each participant performs forward propagation and backward propagation calculation based on its local supply chain business data to generate a local gradient vector; Applying random noise perturbation to the local gradient vector to obtain a perturbed gradient vector, including: According to a preset privacy budget , determine a laplace noise scale parameter : ； for the gradient sensitivity, which is an upper bound on the LI norm of the difference of the gradient vectors produced by any two adjacent data sets; adding, independently for each dimension of the local gradient vector, random noise conforming to a Laplace distribution, forming a perturbed gradient vector; Splitting the perturbed gradient vector into multiple shares according to a threshold secret sharing scheme, and sending each share to other participants using a homomorphic encryption mechanism; After receiving the encrypted gradient shares from all other participants, each participant performs linear superposition on all received shares according to a preset aggregation rule to reconstruct the encrypted aggregation result of the global gradient; Based on the encrypted aggregation result of the global gradient, perform local model parameter update operation to generate the local model required for the next iteration; Repeat the above steps until the model converges or the maximum number of iterations is reached, and finally each participant obtains a consistent global model copy for supply chain prediction, inventory optimization or demand analysis tasks.

2. The federated learning based supply chain data security sharing method according to claim 1, characterized in that, The perturbed gradient vector is split into multiple shares according to a threshold secret sharing scheme, including: Set the total number of participants to , the reconstruction threshold to , where is less than and greater than half of . For each element of the perturbed gradient vector, construct a random polynomial of degree with constant term equal to the value of the element and with the remaining coefficients chosen uniformly at random from the finite field. The value of the polynomial at a non-zero integer point is taken as a share and distributed to other participants.

3. The federated learning based supply chain data security sharing method according to claim 2, characterized in that, The homomorphic encryption mechanism is used to send each share to other participants, including: Each participant generates a pair of Paillier public and private keys in advance and broadcasts the public key to all other participants; The sender encrypts the gradient share using the receiver's public key and transmits it; The receiver decrypts the ciphertext share using its own private key for subsequent aggregation calculation.

4. The federated learning based supply chain data security sharing method according to claim 3, characterized in that, After receiving the encrypted gradient shares from all other participants, each participant performs linear superposition on all received shares according to a preset aggregation rule to reconstruct the encrypted aggregation result of the global gradient, including: Each participant performs element-wise homomorphic summation on all encrypted gradient shares it holds to obtain partial aggregation ciphertext. When at least parties complete share exchange, any parties recover the complete global perturbation gradient vector by Lagrange interpolation formula, which is the arithmetic mean of all parties' local perturbation gradient.

5. The federated learning based supply chain data security sharing method according to claim 4, characterized in that, Based on the encrypted aggregation result of the global gradient, perform local model parameter update operation to generate the local model required for the next iteration, including: Each participant maintains a momentum buffer locally and fuses the reconstructed global perturbed gradient with historical momentum items according to a preset decay factor to generate an update direction; Combine the preset learning rate to perform synchronous adjustment on the current local model parameters.

6. The federated learning based supply chain data security sharing method according to claim 5, characterized in that, The supply chain business data includes purchase order records, logistics transportation status, warehouse inventory level, sales terminal flow and supplier credit rating information; the local machine learning model is a deep neural network whose input layer dimension matches the feature dimension of the supply chain business data, and the output layer is configured as a classification head or a regression head according to the specific task type.

7. The federated learning based supply chain data security sharing method according to claim 6, characterized in that, Before each iteration, each participant verifies and registers the participant identity for the current round through a blockchain consensus mechanism; only participants who have passed the identity verification can join the gradient sharing and aggregation process of the current round; the blockchain network records the participant list, aggregation timestamp and model version hash value of each iteration.

8. The federated learning based supply chain data security sharing method according to claim 7, characterized in that, If a participant has a communication interruption or response timeout during the gradient share transmission process, the remaining participants determine it as a failed node according to a preset fault tolerance strategy; The contribution of the node is automatically excluded in the subsequent aggregation phase, and the reconstruction threshold of the secret sharing is dynamically adjusted to maintain the system robustness; when the number of effective participants is less than the current round is terminated and an abnormal alarm is triggered.

9. The federated learning based supply chain data security sharing method according to claim 8, characterized in that, The privacy budget It is set to a larger value at the beginning of training to ensure the convergence speed of the model, and gradually reduced at the later stage of training to enhance the strength of privacy protection; the dynamic adjustment strategy is written into the smart contract by all participants through the way of consensus in advance, and automatically executed in each iteration.

Images