A randomization-based privacy protection method and system for federated learning of multiple agents

Abstract

This invention discloses a privacy-preserving method and system for multi-agent federated learning based on randomization, relating to the fields of artificial intelligence and distributed computing. The method includes: each client determining its own multiple cooperative groups based on a common random seed; within each cooperative group, each client calculating its own mask vector based on the common random seed and encrypting the model increment so that the aggregator within the group obtains the plaintext aggregation result; calculating the diversity score and weight of each group, and updating the local model based on the weights. All clients make autonomous decisions based on the common random seed, eliminating dependence on a central server. The common random seed dynamically and randomly divides cooperative groups, breaking fixed topologies and promoting information mixing; homomorphic masking technology is used within the groups for secure aggregation, achieving efficient privacy protection.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and distributed computing technology, and in particular to a privacy-preserving method and system for multi-agent federated learning based on randomization. Background Technology

[0002] In the supply chain resource scheduling scenario of ERP (Enterprise Resource Planning) system, each node in the supply chain (such as warehousing, logistics, production, and procurement nodes) accumulates a large amount of sensitive data related to resource scheduling (such as transportation costs, storage capacity, and scheduling plans). This data is the core support for resource scheduling optimization, but at the same time, data privacy must be strictly protected and cannot be shared at will.

[0003] Currently, federated learning technology has been applied to data collaboration and model training for supply chain resource scheduling. However, most of them adopt a centralized federated learning architecture, which is highly dependent on the central server. This architecture not only poses the risk of single point of failure of the central server, but also may lead to privacy leakage throughout the entire chain due to attacks on the central node. It is difficult to adapt to the core requirements of multi-node collaboration and distributed decision-making in ERP supply chain resource scheduling.

[0004] To overcome the limitations of centralized federated learning architectures, decentralized federated learning architectures can eliminate the dependence on a central server, aligning with the multi-node distributed deployment characteristics of ERP supply chain resource scheduling. However, existing decentralized methods still face two core problems when applied to supply chain resource scheduling scenarios: First, there is a lack of efficient and low-overhead trusted verification mechanisms among intelligent agents (i.e., various scheduling nodes in the supply chain), making it impossible to effectively determine whether model parameters and related data have been tampered with. In supply chain resource scheduling, parameter tampering may lead to scheduling decision errors (such as imbalance in transportation capacity allocation and waste of warehousing resources). At the same time, the risk of privacy leakage will increase dramatically with the increase in the number of interactions between intelligent agents, seriously threatening the security of core sensitive data of each node.

[0005] Second, existing intelligent agent systems use fixed communication topologies (such as rings and grids). This rigid collaborative relationship is incompatible with the dynamic nature of ERP supply chain resource scheduling, resulting in slow information dissemination. When updating the model, it is difficult to utilize the fully mixed resource scheduling information globally, and it is easy to get trapped in local optima. This makes the trained model have weak generalization ability and cannot adapt to different regions and different types of resource scheduling needs (such as peak-hour capacity scheduling and cross-node material allocation). Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a privacy-preserving method and system for multi-agent federated learning based on randomization. All clients make autonomous decisions based on a shared random seed, eliminating dependence on a central server. By dynamically and randomly dividing collaborative groups through the shared random seed, fixed topologies are broken, promoting information mixing. Homomorphic masking technology is used within the groups for secure aggregation, achieving efficient privacy protection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a privacy-preserving method for multi-agent federated learning based on randomization, comprising: Each client obtains a common random seed generated by a trusted random source generator, and determines its own multiple collaborative groups based on the common random seed of the current training round; Within each collaborative group, each client calculates its own mask vector based on a common random seed and its group identifier, and encrypts the locally trained model increment based on its own mask vector. The encrypted model increment is then sent to the aggregator within the group so that the aggregator can calculate the plaintext aggregation result. The mask vectors of all clients within each collaborative group satisfy the zero-sum constraint. Based on the mean of the local data feature vectors of all clients within each collaborative group, the diversity score of each group is calculated. The aggregation weight of each group is obtained through the diversity score. After each client receives the plaintext aggregation results of multiple collaborative groups, it updates its local model according to the aggregation weight.

[0008] As an alternative implementation, each client selects a random number from the client set based on a common random seed, a group index, and its own identifier, using a key derivation function and a deterministic pseudo-random number generator. Each client forms a collaboration group; each client repeats this process. Next, get A collaborative group.

[0009] As an alternative implementation, within each collaboration group, each client receives a signature message from other clients in the group to verify each other's identity and data integrity; the signature message is... ;in, For the client The private key is used to generate digital signatures; It is the identifier for the current training round; It is the group identifier to which the client belongs; It is a model increment The Merkle root hash serves as a guarantee of data integrity; It is a statistical feature vector of the local data distribution on the client side; Represents a timestamp.

[0010] As an alternative implementation, the zero-sum constraint is: [Previous...] The mask vector for the client is generated by a deterministic pseudo-random number generator. The mask vector for each client is derived from the previous one. The sum of the negatives of the mask vectors is determined.

[0011] As an alternative implementation, the encrypted model increment is the sum of each client's model increment and its respective mask vector. Each client sends the encrypted model increment, along with a Merkle path proof that proves the encrypted model increment and the root hash of the model increment, to the aggregator within the group. When the aggregator within the group calculates the plaintext aggregation result, the mask vectors of each client cancel each other out.

[0012] As an alternative implementation, the root hash of the model increment is calculated as follows: After local training is completed on each client, the model increment is calculated. The model increment is converted into a deterministic byte sequence through serialization and divided into fixed-size contiguous data blocks. A hash tree is then recursively constructed from bottom to top, and a hash value is calculated for each leaf node. Subsequently, the hash values ​​of internal nodes are calculated, and the calculation structure is distinguished by cascading child node hashes and introducing domain separators. Finally, the hash value at the root of the tree is the digital fingerprint of the model increment.

[0013] As an alternative implementation, the mean of the local data feature vectors of all clients within each collaborative group is used for each group. Feature center: ,in This represents the mean of the feature vectors of all clients within the group. Delegation Group The number of clients in the middle, It is a client The data feature vectors are used to calculate the diversity score of each group through feature centers.

[0014] As an alternative implementation method, the aggregation weights of each group are: ;in, Representative assigned to the first Aggregation results of each group The weights satisfy ; Indicates temperature parameter; It is an exponential function. Score for diversity.

[0015] Secondly, the present invention provides a privacy-preserving system for multi-agent federated learning based on randomization, comprising: The grouping module is configured to obtain a common random seed generated by a trusted random source generator for each client, and determine the multiple cooperative groups to which each client belongs based on the common random seed of the current training round; The intra-group aggregation module is configured such that within each collaborative group, each client calculates its own mask vector based on a common random seed and its group identifier, encrypts the locally trained model increment based on its own mask vector, and sends the encrypted model increment to the intra-group aggregator so that the intra-group aggregator can calculate the plaintext aggregation result; wherein, the mask vectors of all clients within each collaborative group satisfy the zero-sum constraint; The multi-source aggregation module is configured to calculate the diversity score of each group based on the mean of the feature vectors of all local data of all clients within each collaborative group, and obtain the aggregation weight of each group through the diversity score. After each client receives the plaintext aggregation results of multiple collaborative groups, it updates the local model according to the aggregation weight.

[0016] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a privacy-preserving method and system for multi-agent federated learning based on randomization. It utilizes verifiable common randomization as the basis for system coordination, enabling all clients to make autonomous decisions based on a common random seed, thus eliminating dependence on a central server and fundamentally removing single points of failure and performance bottlenecks. By dynamically and randomly dividing collaborative groups through the common random seed, fixed topologies are broken, promoting information mixing. Homomorphic masking is used within each group for secure aggregation, ensuring that the aggregator obtains the correct aggregation result without being able to snoop on the original updates of any individual client, achieving efficient privacy protection. Furthermore, each training round reshuffles the collaborative relationships using a new random seed. This dynamic nature ensures load balancing and fairness in the collaborative network and effectively mitigates model bias and local optima traps caused by non-independent and identically distributed data, thereby improving model generalization ability.

[0018] This invention proposes a privacy-preserving method and system for multi-agent federated learning based on randomization. It addresses the problems of centralized dependence, high privacy leakage risk, and rigid collaboration mechanisms faced by existing federated learning technologies in ERP supply chain resource scheduling scenarios. All ERP supply chain resource scheduling clients (such as warehousing, logistics, and production nodes) autonomously complete local model training, collaborative group construction, and parameter interaction decisions based on random seeds generated from a unified and trusted random source. This eliminates strong dependence on a central server, removes single points of failure and performance bottlenecks, and achieves a balance between high security and high aggregation efficiency while ensuring the privacy of supply chain resource scheduling data. It adapts to the multi-node, distributed business deployment requirements of supply chain resource scheduling. This approach not only achieves secure and privacy-preserving core supply chain scheduling data but also significantly improves the training efficiency, stability, and generalization ability of the resource scheduling model, providing a practical technical solution for the intelligent and secure upgrade of ERP supply chain resource scheduling.

[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 Here is a flowchart of the privacy protection method for multi-agent federated learning based on randomization provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram illustrating the client's information privacy protection principle during an iteration, as provided in Embodiment 1 of the present invention. Figure 3 This is a diagram of a privacy-preserving system architecture based on randomization for multi-agent federated learning, provided in Embodiment 1 of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “including”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Federated learning, as an emerging distributed machine learning paradigm, is based on the core idea of ​​"data does not move, model moves," meaning that each participant trains the model locally without exchanging the original data, thereby achieving collaborative modeling and training while protecting data privacy.

[0027] Traditional federated learning often employs a centralized architecture, relying on a central server for parameter aggregation. However, in a centralized architecture, if the central server is attacked or goes down, the entire system will be paralyzed. At the same time, all traffic converges at the central node, which can easily create communication and processing bottlenecks as the number of agents increases, resulting in poor system scalability.

[0028] To overcome the limitations of centralized federated learning architectures, decentralized federated learning architectures can eliminate the dependence on a central server. However, in existing decentralized methods, the lack of trusted verification between agents makes it impossible to determine whether data has been tampered with, thus the risk of privacy breaches increases dramatically with the number of interactions between agents.

[0029] In addition, the existing federated learning system with fixed communication topology has a rigid collaboration mechanism between agents, which leads to slow information propagation, difficulty in utilizing fully mixed global information when updating the model, easy getting trapped in local optima, and weak generalization ability.

[0030] Therefore, this invention proposes a privacy-preserving method for multi-agent federated learning based on randomization. It utilizes verifiable common randomization as the basis for system coordination, allowing all clients to make autonomous decisions based on a common random seed, thus eliminating dependence on a central server and fundamentally eliminating single points of failure and performance bottlenecks. By dynamically and randomly dividing collaborative groups through the common random seed, fixed topologies are broken, promoting information mixing. Homomorphic masking is used within each group for secure aggregation, ensuring that the aggregator obtains the correct aggregation result without being able to snoop on the original updates of any individual client, achieving efficient privacy protection. Furthermore, each training round reshuffles the collaborative relationships using a new random seed. This dynamism ensures load balancing and fairness in the collaborative network and effectively mitigates model bias and local optima traps caused by non-independent and identically distributed data, improving model generalization ability. This invention not only decentralizes the system but also provides reliable verification for model updates and addresses the rigidity of the collaborative mechanism by reshuffling collaborative relationships in each iteration.

[0031] Example 1 This embodiment provides a privacy-preserving method for multi-agent federated learning based on randomization, including: Each client obtains a common random seed generated by a trusted random source generator, and determines its own multiple collaborative groups based on the common random seed of the current training round; Within each collaborative group, each client calculates its own mask vector based on a common random seed and its group identifier, and encrypts the locally trained model increment based on its own mask vector. The encrypted model increment is then sent to the aggregator within the group so that the aggregator can calculate the plaintext aggregation result. The mask vectors of all clients within each collaborative group satisfy the zero-sum constraint. Based on the mean of the local data feature vectors of all clients within each collaborative group, the diversity score of each group is calculated. The aggregation weight of each group is obtained through the diversity score. After each client receives the plaintext aggregation results of multiple collaborative groups, it updates its local model according to the aggregation weight.

[0032] This embodiment targets resource scheduling in the supply chain management module of an ERP system. It is used for planning the scheduling of supply chain resources (materials, transportation capacity, warehousing space, production equipment, etc.). The client corresponds to each independent participating node in the supply chain resource scheduling in the ERP system, including each warehousing node, logistics carrier terminal, production workshop terminal, and procurement node terminal. Each client only holds its own local data related to resource scheduling and does not share the original data. The aggregator is a trusted node of the ERP system's supply chain resource scheduling management platform (authorized by the enterprise's supply chain management department, only responsible for aggregating resource scheduling-related model parameters, and does not touch the original privacy data).

[0033] The datasets used in this solution are all derived from local business data directly related to resource scheduling at each node of the supply chain in the ERP system. These include warehousing resource data (inventory capacity, available space, inbound and outbound scheduling records), logistics capacity data (capacity scale, transportation costs, scheduling routes, capacity occupancy), production resource data (equipment capacity, equipment occupancy rate, production scheduling records), material scheduling data (material inventory, scheduling path, scheduling time), and historical data of cross-node collaborative scheduling. These data constitute the training dataset.

[0034] Local data on each client is standardized (e.g., unified resource coding, unified scheduling time format, and removal of abnormal scheduling data), invalid data is removed, and sensitive data is anonymized (e.g., capacity costs, material scheduling quotations, and node core resource reserves are anonymized). Local training datasets are categorized according to resource scheduling sub-scenarios (e.g., warehouse resource allocation scenario corresponds to warehouse data + material scheduling data, and capacity scheduling scenario corresponds to capacity data + route data) to ensure that the data is highly matched with the resource scheduling model training task and improve the model scheduling accuracy.

[0035] Each client trains the machine learning model locally based on the aforementioned local training dataset. The machine learning model can employ linear programming models (for material allocation and resource quota optimization), random forest models (for scheduling priority evaluation and resource occupancy prediction), LSTM models (for time series prediction of resource scheduling volume), reinforcement learning models (for dynamic resource scheduling optimization), etc., to reduce deployment difficulty and computational cost, and ensure the real-time nature of scheduling decisions.

[0036] During local training, the objective function of the resource scheduling scenario is combined (with the core objectives of maximizing resource utilization, minimizing scheduling costs, and minimizing scheduling time, while also taking into account scheduling priorities). The model parameter initialization is uniformly configured by the ERP supply chain resource scheduling management platform to ensure that the model initialization of each client is consistent, improve aggregation efficiency, and meet the global optimization needs of resource scheduling.

[0037] The model parameters include feature parameters related to resource scheduling (warehouse available space coefficient and material scheduling priority coefficient in the warehouse resource allocation model, and transportation capacity utilization rate and scheduling cost coefficient in the transportation capacity scheduling model). The parameter dimensions are aligned with the business needs of resource scheduling to avoid redundant parameters. The incremental calculation of model parameters must be based on local resource scheduling data to ensure the effectiveness and relevance of the incremental data and support the optimization of scheduling decisions.

[0038] The trained model is deployed in the supply chain resource scheduling module of the ERP system. Each client model is only used for resource scheduling decisions of its own node (local warehousing resource allocation, transportation capacity scheduling planning). The aggregated global model is used for overall optimization of cross-node resource scheduling in the supply chain (global material collaborative scheduling, cross-regional transportation capacity allocation). Model updates need to be synchronized to the ERP system to ensure linkage with the resource scheduling business process and realize real-time updates of scheduling decisions.

[0039] Based on the above limitations, the following is combined with Figures 1-2 The method of this embodiment will be described in detail below.

[0040] Step 1: Local training and preparation process.

[0041] Each client (each node of ERP supply chain resource scheduling, including warehousing nodes, logistics carriers, and production workshops) performs local initialization and training of the model based on its own preprocessed local resource scheduling training dataset (including warehousing resource data, logistics capacity data, production resource data, material scheduling data, and cross-node collaborative scheduling historical data) using a linear programming model.

[0042] Specifically, this process can be precisely described as a parameter update function: ; in, Represents the local training function; Indicates client In the The global model parameters obtained after a round of training are a high-dimensional real vector, and their dimension is determined by the selected neural network architecture. These are the latest model parameters obtained after local training; Representing the client The local training dataset, specifically in the form of , which includes One input / output pair; It is the local learning rate, which controls the step size of parameter updates; It is a positive integer representing the number of rounds of local training, that is, the number of times the local dataset is fully traversed, which is used to adjust the training depth of each client on the local training dataset; This represents the loss function, used to measure the accuracy of the model's predictions.

[0043] Furthermore, from the perspective of optimization theory, the above local training process is essentially solving a regularized empirical risk minimization problem: ; Among them, the second item These are proximal terms, and their purpose is to prevent local updates from deviating excessively from the previous round's global model. It is the corresponding regularization coefficient; For the first One input / output pair.

[0044] In practice, this is typically achieved through multiple rounds of mini-batch stochastic gradient descent. In each round of training, the local training dataset is first randomly shuffled and then divided into several mini-batches. For each mini-batch, the gradient of the loss function with respect to the model parameters is calculated. The model parameters are then updated along the negative gradient direction: Among these, the learning rate and batch size are key hyperparameters, which together affect the stability and convergence speed of training.

[0045] Furthermore, after local training is complete, the client calculates the model increment. This refers to the parameter update based on the resource scheduling objective function, including parameter adjustments related to scheduling cost optimization and resource utilization improvement; this is a vector difference operation, model increment. It is a vector with the same dimension as the original model parameters, and each component... This represents the amount of change in the corresponding parameter.

[0046] Furthermore, to ensure model increment To ensure integrity during transmission and aggregation, the client generates a verifiable cryptographic commitment by constructing a Merkle tree and obtaining its root hash. To achieve this.

[0047] The specific process is as follows: First, increment the model The data is converted into a deterministic byte sequence through serialization and then divided into fixed-size contiguous blocks of data. ; Then, a hash tree is recursively constructed from bottom to top, and the hash value is calculated for each leaf node. , where the field separator With block index Together, they ensured the uniqueness and fixed position of the leaf nodes. For cryptographic hash functions; Then calculate the internal nodes. It hashes child nodes and introduces a domain separator. To distinguish computational structures; Ultimately, the hash value at the root of the tree It is defined as a unique digital fingerprint of the entire model increment.

[0048] It should be noted that this root hash has explicit cryptographic binding properties, meaning that due to the collision resistance and avalanche effect of the hash function, it is cryptographically bound to the target hash. Any local modifications will be propagated through the tree structure and result in The unpredictable changes in the root hash allow any participant to efficiently verify the integrity of the received data by reconstructing and comparing the root hash, thus achieving strong data integrity guarantees with minimal overhead in a distributed environment.

[0049] Furthermore, each client accesses the local dataset. Extracting statistical feature vectors This is used to quantify and describe the distribution characteristics of a local dataset, including statistical features such as the proportion of class samples, the entropy of the label distribution, and the mean and variance of the input features.

[0050] Feature vector It plays an important role in subsequent collaborative computing, and is used to assess the diversity of data distribution among different client groups, the weights when computing model aggregation, and to help detect possible data anomalies or malicious behavior.

[0051] Step 2: Dynamic Group Construction. The goal is to enable all clients to independently compute their own multiple cooperative groups based on a common random seed, without the need for central scheduling. Each cooperative group is used for subsequent secure aggregation. Through this multi-home random grouping, each client can participate in the aggregation process of multiple groups simultaneously, laying the foundation for subsequent adaptive fusion.

[0052] Specifically: Let the set of all clients be ,in This indicates the total number of active clients in the system. Representing the A unique identifier for each client.

[0053] In the In the round, the system obtains a common random seed from a verifiable random beacon. .

[0054] Each client belong There are 1 collaborative groups, among which This is a preset hyperparameter representing the number of groups each client participates in. Client Belonging to the One collaborative group , It is generated using the following deterministic algorithm: First, using a seed Group Index and client-side identifier Generate a group-specific random seed ,in For key derivation functions (such as HKDF or PBKDF2), " is a fixed field separator used to distinguish random number generation for different purposes; then, with For input, a deterministic pseudo-random number generator (such as AES-CTR or ChaCha20) and a sampling algorithm are used to obtain the data from the client set. Selected from A group of clients constitutes a collaboration group. ,in This is the preset group size.

[0055] Furthermore, the sampling algorithm can employ one of the following two methods: random permutation, i.e., using... Perform a deterministic random permutation of the list of client identifiers (e.g., Fisher-Yates shuffle), and take the first few results of the permutation. Each element is used as a group member; independent sampling method: utilizing As a seed, a deterministic random number generator is repeatedly used to draw random numbers from the set without replacement. One client.

[0056] Since the above process is entirely based on common input The decision is made that all clients will use the same... Calculated groups They have a completely consistent membership composition.

[0057] Furthermore, after determining group affiliation, each client needs to exchange metadata with group members and verify each other's identities and data integrity.

[0058] Specifically, the client Generate and broadcast a signed message: ; in, Representing the client The private key is used to generate digital signatures; It is the identifier for the current training round; It is the group identifier to which the client belongs; It is a model increment The Merkle root hash serves as a guarantee of data integrity; It is a statistical feature vector of the local data distribution on the client side; This represents a timestamp, used to ensure the freshness of the message.

[0059] When a client receives a message from other members of the group Signature message When using the corresponding public key Execute the verification function ,in Representing the client Metadata messages broadcast for it The digital signature created.

[0060] This verification process is used to verify and ensure that the message does indeed come from the claimed sender. The message was not tampered with during transmission and the sender does indeed belong to the claimed group. .

[0061] Step 3: Secure Aggregation. This is the core step in achieving privacy protection. Its purpose is to correctly aggregate model updates within a group without exposing any individual client model updates.

[0062] This process ensures that the aggregator can calculate the correct aggregation result without knowing the original data of any participant, i.e., it can only calculate the sum of the increments of all client models within the group. without exposing individual model increments for each client. .

[0063] Specifically, homomorphic masking is used to achieve this goal.

[0064] Set up a collaborative group Include Client The model increment for each client is Its dimensions are All clients use a common random seed. and group identifier Independently calculate the mask vector , .

[0065] Furthermore, the mask vector satisfies the zero-sum constraint. .

[0066] Specifically, before The mask for each client is generated by a deterministic pseudo-random number generator, that is: ; No. The mask for each client is changed from the previous one. The negative sum of the masks is determined, that is ; in, This refers to a cryptographically secure pseudo-random number generator, such as one employing the AES or ChaCha20 algorithm; It is a field separator to prevent random number reuse attacks; It is a group identifier that ensures that the masks of different groups are independent; It is the index of the client within the group, ensuring that the masks of different clients within the group are independent.

[0067] Furthermore, each client Use its proprietary mask Model increment Encryption steps ,in It is the encrypted model increment, which appears as random noise.

[0068] Furthermore, after encryption is complete, the client will... Send to the aggregator within the group, where It is proof Compared to the previously promised root hash The corresponding Merkle path proof.

[0069] Furthermore, after the aggregator collects the encrypted updates from all group members, it first verifies each one. The Merkle proof ensures data integrity. After successful verification, the aggregator calculates a weighted sum of the encrypted updates: .

[0070] Due to zero-sum constraint It can be known that .

[0071] The above equation shows that, although the aggregator has always been processing encrypted data But what was ultimately obtained It is exactly equal to the sum of the original model increments, and the aggregator has never come into contact with any plaintext. However, the correct aggregation result was obtained.

[0072] Meanwhile, each client uses a common random seed to generate its own zero-sum homomorphic mask. When the aggregator sums the results, it ensures that the mask can cancel each other out and does not reveal parameters such as resource scheduling costs and resource reserves.

[0073] Furthermore, from an information theory perspective, the strength of privacy protection provided by this step depends on the randomness of the mask. For the aggregator or any external observer, knowing only... In this case, to deduce the original This is equivalent to the equation Solving in the middle And because It is uniformly and randomly generated from a sufficiently large random space, and the attacker does not know the specifics. The value, therefore, for a single client, allows the attacker to successfully guess... The probability is ,in Represents the cardinality of the mask space.

[0074] It should be noted that, for a Given a dimensional vector, the probability of correctly guessing all its components is... For typical deep learning models This probability is negligible.

[0075] Step 4: Adaptive Intelligent Weight Update. This is key to improving the model's generalization ability. Its goal is to assign appropriate weights to the aggregation results from different sources based on the diversity of data distribution in each co-group, thereby better utilizing global information in model updates.

[0076] Specifically: Set up the client At the same time belong to One collaborative group and received accordingly Aggregation results .

[0077] To calculate the diversity score for each group, we first calculate the diversity score for each group. Feature center ,in This represents the mean of all client feature vectors within the group, signifying the average data distribution characteristics of that group. Delegation Group The number of clients in the middle, It is a client The data feature vector.

[0078] Furthermore, based on the feature centers of each group, various diversity measurement methods can be adopted.

[0079] For example, first calculate the group covariance matrix Then you can get the group The diversity score is ,in Represents the covariance matrix The eigenvalues ​​of the covariance matrix are essentially the geometric mean of the determinants of the covariance matrix, which can reflect the scatter volume of the data distribution in the feature space.

[0080] Alternatively, similarity-based methods can be used: ; The summation term calculates the average similarity of all client pairs within the group, and the cosine similarity is defined as... , and These are the L2 norms of the vectors. In the above formula, the final diversity score is 1 minus the average similarity, to ensure that the lower the similarity (the greater the difference), the higher the diversity score.

[0081] Furthermore, since the heterogeneity of data distribution in different collaborative groups contributes unevenly to the model's generalization ability, weights are calculated using diversity scores.

[0082] Specifically, after obtaining the diversity scores for each group, the scores are converted into fusion weights using a softmax function with a temperature parameter: ; in, Representative assigned to the first Aggregation results of each group The weights that satisfy ; This represents a temperature parameter used to control the sharpness of the weight distribution; It is an exponential function used to ensure that the weights are always positive; L is the total number of groups.

[0083] It should be noted that when the temperature parameter At that time, all weights tend to be equal, that is This is equivalent to a simple average; when At this time, only the group with the most diversity receives a weight of 1, while the other groups receive a weight of 0, which is equivalent to selecting the optimal group; typically, the temperature parameter can be set to... This is to strike a balance between exploration (considering all groups) and utilization (focusing on highly diverse groups).

[0084] Furthermore, the final model update is achieved through the following weighted fusion: ; in, It is the first The model parameters of the previous round for each client; It is the global learning rate, used to control the overall update step size; It is a weighted average of the aggregation results.

[0085] Furthermore, it is worth emphasizing that the above adaptive weight calculation guides the model toward a more generalizable solution space region by emphasizing the update direction of the highly diverse group.

[0086] Finally, once the model converges and training is complete, an updated resource scheduling model is obtained and synchronized to the supply chain resource scheduling module of the ERP system for local resource scheduling decisions (dynamic allocation of warehousing resources and real-time scheduling of transportation capacity). The global aggregation model is used for overall optimization of cross-node resource scheduling in the supply chain (global material allocation and cross-regional transportation capacity collaboration), improving the efficiency and rationality of resource scheduling.

[0087] In this embodiment, the above scheme is implemented using a three-tier distributed architecture, such as... Figure 3 As shown, the components at each layer work together, specifically including: The client-side intelligent agent layer consists of multiple client-side intelligent agents, each client... It is a fully functional autonomous unit that includes a local model manager for storing current and historical versions, a privacy-preserving computing engine for performing homomorphic mask generation, digital signatures and verification, and a data feature analyzer for calculating and generating local data distributions.

[0088] The Trusted Random Source Layer contains a Trusted Random Source Generator, which provides an unbiased, verifiable, and unpredictable common random number seed.

[0089] To adapt to different deployment environments and security requirements, there are two methods for generating random sources: (1) A distributed random beacon based on VRF, in which each node submits VRF output, and a common seed is generated by aggregation, and any node can verify its correctness; (2) Threshold signature-based random beacons, which generate a unique signature through cooperation among multiple nodes, using the hash value of the signature as a common seed.

[0090] It is worth noting that the VRF scheme represents the design philosophy of maximum decentralization, while the threshold signature scheme embodies the engineering thinking of prioritizing practicality and efficiency, and can be deployed in combination according to existing needs in practice.

[0091] The peer-to-peer communication network layer uses a P2P network topology (such as the Gossip protocol or DHT) to build an overlay network, which is responsible for the transmission of all control information and model update data between agents.

[0092] To address the core technical pain points of existing federated learning technologies in ERP supply chain resource scheduling scenarios, such as centralized dependence, high privacy leakage risks, rigid collaboration mechanisms, and insufficient model generalization capabilities, this embodiment proposes a privacy-preserving method based on randomization-based multi-agent federated learning. Through a collaborative randomization mechanism and dynamic cooperation architecture, it achieves the following significant benefits: 1. Eliminate centralized dependence, remove single points of failure and performance bottlenecks in the system, and adapt to distributed scheduling scenarios.

[0093] Using verifiable shared randomization as the core foundation for global system coordination, all ERP supply chain resource scheduling clients (such as warehousing, logistics, and production nodes) autonomously complete local model training, collaborative group construction, and parameter interaction decisions based on random seeds generated from a unified and trusted random source, eliminating strong dependence on a central server. Compared to the shortcomings of traditional centralized federated learning architectures, where the central node is prone to becoming a performance bottleneck (unable to support large-scale concurrent collaboration of supply chain nodes) and a single point of failure (central node failure leading to end-to-end scheduling interruption), the decentralized architecture of this method allows the ERP supply chain resource scheduling system to flexibly expand the number of nodes (such as adding regional warehouses or new logistics partners), with each client autonomously participating in collaboration. This fundamentally ensures the high availability, high scalability, and stability of the supply chain resource scheduling system, perfectly adapting to the multi-node, distributed business deployment requirements of supply chain resource scheduling.

[0094] 2. Dynamic collaboration + secure aggregation enables efficient privacy protection and data security protection for scheduling.

[0095] Based on a shared random seed, the system dynamically and randomly divides supply chain resource scheduling collaboration groups, breaking the rigidity of fixed topologies (rings, grids) in traditional intelligent agent systems. This allows for real-time aggregation of nodes with high resource scheduling correlation (such as warehousing and logistics nodes in the same area, or production and procurement nodes for the same material), promoting the full mixing and flow of core supply chain scheduling information (inventory capacity, transportation costs, equipment occupancy, etc.). Simultaneously, a homomorphic mask secure aggregation technology is employed within the collaboration groups: each client generates its own zero-sum homomorphic mask to encrypt model parameter updates using a random seed. The aggregator can only receive encrypted parameters, calculates the weighted sum, and then uses the mask to cancel out the plaintext aggregation result, ensuring that no single client's original parameters or sensitive scheduling data can be accessed throughout the process. This design achieves a balance between high security and high aggregation efficiency while ensuring the privacy of supply chain resource scheduling data (such as core warehousing reserves, logistics quotations, and production capacity). It addresses the industry pain point of privacy leaks increasing dramatically with the number of interactions in traditional decentralized methods, meeting the compliance and security requirements for ERP supply chain resource scheduling data privacy protection.

[0096] 3. Dynamically shuffling collaborative relationships solves the problem of non-independent and identically distributed data, improving the model's generalization ability and scheduling accuracy.

[0097] In each iteration of model training, the collaborative relationships among supply chain resource scheduling clients are reshuffled using a new random seed, rather than a fixed collaborative topology. This dynamism brings three core values: Ensure load balancing in the collaborative network: avoid long-term overload of core scheduling nodes (such as major logistics providers and core warehouses) and insufficient participation of ordinary nodes (such as small production workshops and regional distribution points), ensure reasonable allocation of resources (computing and storage) for each node, and improve the overall collaborative efficiency of the system. Ensuring fairness in collaboration: Avoiding information barriers between nodes caused by fixed topology, ensuring that resource scheduling data from each client (such as inventory data from different regions and equipment data from different production lines) can fully participate in global model training, eliminating data silos; Effectively mitigate model bias caused by non-independent and identically distributed (Non-IID) data: Resource scheduling data at each node of the supply chain (such as logistics data during peak and off-peak seasons, and warehousing data in different regions) naturally have different distributions. Dynamic collaboration allows the model to fully learn the scheduling characteristics of different nodes, avoid getting trapped in local optima, and ultimately improve the generalization ability of the global model to different scenarios of ERP supply chain resource scheduling (such as peak-hour capacity scheduling and cross-regional material allocation), making the optimized resource scheduling scheme more in line with actual business needs, reducing scheduling costs, and improving resource utilization.

[0098] In summary, the method in this embodiment, through the triple innovation of decentralized architecture, dynamic privacy aggregation, and dynamic collaboration mechanism, adapts to the business characteristics and technical pain points of ERP supply chain resource scheduling. It not only achieves the security and privacy protection of core supply chain scheduling data, but also significantly improves the training efficiency, stability, and generalization ability of the resource scheduling model, providing a practical and feasible technical solution for the intelligent and secure upgrade of ERP supply chain resource scheduling.

[0099] It should be noted that all data acquisition is conducted in accordance with laws and regulations and with user consent, and the data is used legally.

[0100] Example 2 This embodiment provides a privacy-preserving system for multi-agent federated learning based on randomization, including: The grouping module is configured to obtain a common random seed generated by a trusted random source generator for each client, and determine the multiple cooperative groups to which each client belongs based on the common random seed of the current training round; The intra-group aggregation module is configured such that within each collaborative group, each client calculates its own mask vector based on a common random seed and its group identifier, encrypts the locally trained model increment based on its own mask vector, and sends the encrypted model increment to the intra-group aggregator so that the intra-group aggregator can calculate the plaintext aggregation result; wherein, the mask vectors of all clients within each collaborative group satisfy the zero-sum constraint; The multi-source aggregation module is configured to calculate the diversity score of each group based on the mean of the feature vectors of all local data of all clients within each collaborative group, and obtain the aggregation weight of each group through the diversity score. After each client receives the plaintext aggregation results of multiple collaborative groups, it updates the local model according to the aggregation weight.

[0101] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0102] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0103] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0104] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0105] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0106] The method in Example 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0107] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0108] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0109] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0110] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0111] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment 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 invention.

[0112] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A privacy protection method based on randomized multi-agent federated learning, characterized in that, include: Each client obtains a common random seed generated by a trusted random source generator, and determines its own multiple collaborative groups based on the common random seed of the current training round; Within each collaborative group, each client calculates its own mask vector based on a common random seed and its group identifier, and encrypts the locally trained model increment based on its own mask vector. The encrypted model increment is then sent to the aggregator within the group so that the aggregator can calculate the plaintext aggregation result. The mask vectors of all clients within each collaborative group satisfy the zero-sum constraint. Based on the mean of the local data feature vectors of all clients within each collaborative group, the diversity score of each group is calculated. The aggregation weight of each group is obtained through the diversity score. After each client receives the plaintext aggregation results of multiple collaborative groups, it updates its local model according to the aggregation weight.

2. The privacy-preserving method for multi-agent federated learning based on randomization as described in claim 1, characterized in that, Each client selects a subset of the set of clients by a key derivation function and a deterministic pseudo-random number generator according to a common random seed, a group index and its own identity A collaborative group is formed by Each client repeats the process Collaborative groups.

3. The randomization-based federated learning privacy protection method for multiple agents according to claim 1, wherein, Within each collaboration group, each client receives a signed message from other clients in the group to verify each other's identity and data integrity; the signed message is... ;in, For the client The private key is used to generate digital signatures; It is the identifier for the current training round; It is the group identifier to which the client belongs; It is a model increment The Merkle root hash serves as a guarantee of data integrity; It is a statistical feature vector of the local data distribution on the client side; Represents a timestamp.

4. The randomization-based federated learning privacy protection method for multiple agents according to claim 1, wherein, The zero-sum constraint is: before The mask vector for the client is generated by a deterministic pseudo-random number generator. The mask vector for each client is derived from the previous one. The sum of the negatives of the mask vectors is determined.

5. A privacy-preserving method for multi-agent federated learning based on randomization as described in claim 1, characterized in that, The encrypted model increment is the sum of each client's model increment and its respective mask vector. Each client sends the encrypted model increment, along with a Merkle path proof that the encrypted model increment corresponds to the root hash of the model increment, to the aggregator within the group. When the aggregator within the group calculates the plaintext aggregation result, the mask vectors of each client cancel each other out.

6. The privacy-preserving method for multi-agent federated learning based on randomization as described in claim 5, characterized in that, The calculation process of the root hash of the model increment is as follows: After the local training of each client is completed, the model increment is calculated. The model increment is converted into a deterministic byte sequence through serialization and divided into fixed-size continuous data blocks. A hash tree is then recursively constructed from bottom to top, and a hash value is calculated for each leaf node. Subsequently, the hash values ​​of the internal nodes are calculated. The calculation structure is distinguished by cascading child node hashes and introducing a domain separator. Finally, the hash value at the root of the tree is the digital fingerprint of the model increment.

7. The randomization-based federated learning privacy protection method for multiple agents according to claim 1, wherein, The mean of the feature vectors of all client-local data within each collaborative group is the mean of each group. Feature center: ,in This represents the mean of the feature vectors of all clients within the group. Delegation Group The number of clients in the middle, It is a client Data feature vectors; Therefore, the diversity score of each group is calculated using feature centers.

8. The randomization-based federated learning privacy protection method for multiple agents according to claim 1, wherein, The aggregate weights for each group are: ;in, Representative assigned to the first Aggregation results of each group The weights satisfy ; Indicates temperature parameter; It is an exponential function. Score for diversity.

9. A randomization-based multi-agent federated learning privacy protection system, characterized in that, include: The grouping module is configured to obtain a common random seed generated by a trusted random source generator for each client, and determine the multiple cooperative groups to which each client belongs based on the common random seed of the current training round; The intra-group aggregation module is configured such that within each collaborative group, each client calculates its own mask vector based on a common random seed and its group identifier, encrypts the locally trained model increment based on its own mask vector, and sends the encrypted model increment to the intra-group aggregator so that the intra-group aggregator can calculate the plaintext aggregation result; wherein, the mask vectors of all clients within each collaborative group satisfy the zero-sum constraint; The multi-source aggregation module is configured to calculate the diversity score of each group based on the mean of the feature vectors of all local data of all clients within each collaborative group, and obtain the aggregation weight of each group through the diversity score. After each client receives the plaintext aggregation results of multiple collaborative groups, it updates the local model according to the aggregation weight.

10. An electronic device, comprising: It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-8.

Images