A mechanism knowledge guided security multi-party collaboration model optimization method and system

Abstract

The application discloses a security multi-party collaborative model optimization method and system under mechanism knowledge guidance, and the method firstly constructs a federal learning environment, including a server S and N a plurality of clients, the server S initializes global model parameters and an initial population position vector.Secondly, each client downloads the initial global model parameters for federal learning, and the server S calculates an aggregation interval of the current communication round based on the parameters uploaded by the clients and the initial population position vector.Then, based on the aggregation interval, the server S controls each client to perform local model training and aggregation, uses a population optimization algorithm to update the population position vector.Finally, the above operation is repeated until a training termination condition is met, and an optimized global model is output, and collaborative model optimization is completed.The application significantly reduces communication overhead under the premise of ensuring model performance, and solves the problem that hyperparameter tuning relies on artificial experience.

Description

Technical Field

[0001] This invention relates to the field of distributed machine learning and artificial intelligence optimization technology, and in particular to a secure multi-party collaborative model optimization method and system guided by mechanistic knowledge. Background Technology

[0002] In recent years, with the large-scale deployment of IoT terminals and edge computing devices, federated learning, as a distributed machine learning paradigm that enables multi-party collaborative modeling without sharing raw data, has been widely used in privacy-sensitive scenarios such as smart healthcare, industrial predictive maintenance, and intelligent transportation. It effectively alleviates the problem of data silos while protecting data privacy.

[0003] However, in real-world industrial scenarios, federated learning systems typically face the following challenges: First, communication networks are subject to jitter and instability, with significant differences in the online frequency and bandwidth conditions of different clients; second, the local data distribution of each client exhibits obvious non-independent and identically distributed characteristics, which can easily lead to slow convergence or even divergence of the global model; and third, the computing power and storage resources of edge devices are limited, making it difficult to support high-frequency model aggregation and complex hyperparameter tuning processes.

[0004] To address the aforementioned issues, most existing federated learning methods employ fixed aggregation frequencies or simple rules to adjust aggregation intervals, failing to effectively integrate the theoretical convergence mechanism of the model with real-time training feedback information, which can easily lead to communication redundancy or local model drift. Key hyperparameters such as learning rate and batch size have high dimensionality and strong coupling, and existing methods mainly rely on manual experience or static search strategies, making it difficult to obtain the global optimal solution.

[0005] Given the aforementioned shortcomings of existing technologies, there is an urgent need to develop a federated learning collaborative optimization method that can integrate theoretical mechanisms and practical feedback and automatically find the best approach. This method can significantly reduce communication overhead while ensuring model performance, thereby improving the engineering applicability of federated learning in real-world complex scenarios. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a method and system for optimizing a secure multi-party collaborative model guided by mechanistic knowledge. By introducing the theoretical convergence boundary and real-time loss feedback of the evidence theory fusion model, it achieves adaptive dynamic adjustment of the aggregation interval and utilizes an improved swarm intelligence algorithm to automate hyperparameter optimization. This solves the problems of low communication efficiency, rigid aggregation strategies, and difficulty in hyperparameter tuning in existing federated learning.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In one aspect, this invention provides a method for optimizing a secure multi-party collaborative model guided by mechanistic knowledge, applicable to a system including a parameter server. S andN A federated learning system for multiple clients, the method comprising the following steps: S1. Construct a federated learning environment, including server S and N clients. Server S and N clients establish an end-to-end encrypted communication channel through the key system provided by KGC (Key Generation Center). Server S initializes global model parameters and combines the hyperparameters to be optimized in the federated learning (learning rate). Sliding window size Fusion decision threshold Activation function slope parameter The encoding is used as the initial population position vector in the population optimization algorithm. .

[0008] S2. Each client downloads the global model parameters from the parameter server S, loads them locally for training, and uploads the local model parameters and loss values. Based on the model parameters and loss values ​​uploaded by the clients, and combined with the initial population position vector, the server S calculates the theoretical convergence boundary as theoretical mechanism information and calculates the global loss value as practical feedback information. The server S uses the Dempster-Schaffer evidence theory (DS) to fuse the theoretical mechanism information and practical feedback information, and determines whether to adjust the aggregation interval based on the fusion result and the size of the fusion decision threshold, and outputs the adaptive model aggregation interval for the current communication round.

[0009] S3. Based on the adaptive model aggregation interval τ obtained in step S2, server S controls each client to perform multiple rounds of local model training and aggregation. After completing the preset number of rounds of federated learning training, the fitness value of the current hyperparameter combination is calculated, and the population optimization algorithm is used to update the individual position vector to generate the next generation of hyperparameter combinations and update the population position vector.

[0010] S4. Repeat steps S2 and S3 until the training termination condition is met, output the optimized global model, and complete the collaborative model optimization.

[0011] Preferably, the theoretical mechanism information in step S2 is the theoretical convergence boundary of the federated learning model, and its specific calculation process includes: Client Calculate the Lipschitz constant of the local loss function based on the local loss value and local model parameters. With Smooth constant And uploaded to the server S ;server S Aggregate constants uploaded by all clients and The Lipschitz constant of the global loss function is obtained. With smoothness constant According to Lipschitz constant With smoothness constant Learning rate Total number of local training sessions and the current aggregation interval The theoretical convergence boundary is calculated based on the convergence upper bound formula of federated learning theory.

[0012] Preferably, the effective feedback information in step S2 is the global model loss value, and its specific calculation process includes: After the client completes local training, it calculates the loss value of its local model and uploads it to the server. S ;server S The received local loss values ​​are weighted and averaged based on the amount of data from each client to obtain the global loss value.

[0013] Step S2 employs the Dempster-Schafer evidence theory to fuse theoretical mechanism information with practical feedback information, specifically including: S21. To address the decision-making requirements for adaptive aggregation frequency in federated learning, the server... S The server aggregates the local model parameters uploaded by each client to obtain the aggregated global model. S The convergence state of the aggregated global model is defined as the recognition framework. ,in This indicates a non-converged state, and the aggregation interval should be maintained in this case. This indicates a convergence state, at which point the aggregation interval should be reduced.

[0014] S22, Server S For the global loss value sequence consisting of several consecutive rounds of global loss values ​​and the theoretical convergence boundary sequence, a sliding window of length Length is introduced for smoothing. The smoothed loss sequence and the theoretical boundary sequence are calculated using a moving average method. The number of times the current sequence value fails to reach the historical minimum value is calculated by comparing the current sequence value with the historical minimum sequence value. When the current sequence value is less than the historical minimum sequence value, the number of times... If the value is 0, then the previous count is used; otherwise, the previous count is used. The value is increased by 1.

[0015] S23, Combine the two numbers As input, the two are used to activate the function. These are respectively mapped to the basic probability assignments in evidence theory. and .

[0016] S24. Using the Jaeger combination rule to... and Fusion yields joint evidence .

[0017] S25, Server S Based on joint evidence With fusion decision threshold Determine whether to adjust the aggregation interval, and send the adjusted aggregation interval to the client, specifically: like If the current state is considered convergent, the aggregation interval will be adjusted. Adjusted to Otherwise, maintain the current aggregation interval. .

[0018] Preferably, the population optimization algorithm used in step S3 is an improved particle population optimization algorithm, specifically including: S31. Server S initializes the population size, population position vector, and velocity.

[0019] S32. Based on fitness values, the population particles are divided into three roles: explorers, followers, and vigilants. For explorer particles, a velocity update formula with random perturbation is used to update their velocity and position, enabling global exploration of the search space. For follower particles, a particle swarm optimization formula is used to update their velocity and position, ensuring overall search efficiency. For vigilant particles, simulated annealing perturbation is introduced based on their current position, and the acceptance of solutions is judged according to the Metropolis criterion. Controlled random jumps maintain population diversity and improve the global optimization capability of hyperparameter search.

[0020] S33. Calculate the fitness value of each particle and update the individual optimal position and the global optimal position.

[0021] S34. Repeat steps S32 and S33 until the training termination condition is met, and output the optimal result.

[0022] Preferably, the individuals in the population optimization algorithm used in step S3 include a global model composed of population position vectors, and the accuracy is tested based on the global model aggregated on the server side under the current population position vectors. The fitness value f is calculated as follows: minus The value, This represents the current number of local training iterations on the client side. These are the weighting coefficients.

[0023] In another aspect, the present invention provides a secure multi-party collaborative model optimization system guided by mechanistic knowledge, comprising the following modules: The Federated Learning Environment Building Module is used to build a federated learning environment, including servers. S and N One client, server S initializes global model parameters and initial population location vector; The aggregation interval calculation module is used by each client to download the initial global model parameters for federated learning. The server S calculates the aggregation interval for the current communication round based on the parameters uploaded by the client and the initial population position vector. Population location vector update module, based on aggregation interval ,server S Control each client to perform local model training and aggregation, use a population optimization algorithm to update the population position vector until the training termination condition is met, output the optimized global model, and complete the collaborative model optimization.

[0024] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding claims.

[0025] Compared with the prior art, the beneficial effects of the present invention are reflected in: 1. This invention achieves dynamic adaptive adjustment of the aggregation interval by fusing evidence from theoretical convergence boundaries and real-time training loss, which significantly reduces communication overhead while ensuring model performance.

[0026] 2. This invention introduces an improved swarm optimization algorithm to automatically find the optimal hyperparameters for federated learning, thus solving the problem that hyperparameter tuning relies on human experience. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a logical block diagram of the DS evidence theory fusion mechanism in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the role division and update strategy of the improved granular swarm algorithm in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the system architecture and hardware deployment of Embodiment 2 of the present invention; Figure 5 This is a graph showing the change in the polymerization interval of the method in Embodiment 1 of the present invention; Figure 6 This is a graph showing the change in accuracy of the method in Embodiment 1 of the present invention. Detailed Implementation

[0028] To make the technical solution, innovative features, and implementation effects of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the following description is merely a typical implementation example of this invention and is not intended to limit the invention. Any modifications or adjustments made based on the principles of this invention should be included within the scope of protection of this invention.

[0029] The accompanying drawings are for illustrative purposes only, and their dimensions, proportions, and other details are not intended to limit practical application. Those skilled in the art can make appropriate adjustments based on actual application needs. The technical solution and implementation process of the present invention will be further described below with reference to specific embodiments.

[0030] Example 1: This invention provides a method for optimizing a secure multi-party collaborative model guided by mechanistic knowledge. This embodiment uses a parameter server. S and N A federated learning system with 1 client, where the client's local dataset is as follows: Client The model parameters are Client The loss function is defined as .like Figure 1 The efficient multi-party collaborative model optimization method shown, which integrates mechanistic knowledge and practical feedback, includes the following steps: S1. Construct a federated learning environment, including servers. S and N One client, one server S Initialize global model parameters and initial population position vector.

[0031] Step S1 specifically involves: constructing a federated learning environment, including servers. S and N One client, one server S and N Each client establishes an end-to-end encrypted communication channel through the key system provided by the Key Generation Center (KGC). S Initialize global model parameters, and the server S The combination of hyperparameters to be optimized in federated learning is encoded into the initial population position vector in the population optimization algorithm. The hyperparameter combination to be optimized includes the learning rate. Sliding window size Fusion decision threshold Activation function slope parameter .

[0032] S2. Each client downloads the initial global model parameters for federated learning, and the server... S Based on the parameters uploaded by the client and the initial population position vector, calculate the aggregation interval of the current communication round; like Figure 2As shown, the reason for introducing the Dempster-Schafer evidence theory in step S2 is that in a real federated learning environment, relying solely on theoretical mechanism information leads to a mismatch between theoretical assumptions and actual scenarios, failing to accurately reflect the model's true convergence state; relying solely on real-time feedback information is susceptible to instantaneous data noise, resulting in distorted decisions and potentially causing invalid aggregation or aggregation delays. Figure 2 The fusion mechanism in the model integrates two pieces of evidence: theoretical fusion information and practical feedback information, which can more accurately determine whether the aggregation interval needs to be adjusted.

[0033] Step S2 includes: S21, Client Calculate the Lipschitz constant of the local loss function based on the local loss value and local model parameters. With smoothness constant Upload to server S The specific calculation formula is as follows: In the formula, Indicates the first The local loss function for each client; Indicates the first The model parameters of each client at the current moment; Represents the global model parameters at the current moment; Indicates the client The gradient of the local loss function.

[0034] server S The received parameters are aggregated based on the number of client samples to obtain the global loss function, Lipschitz constant. With smoothness constant And calculate parameters The specific calculation formula is as follows: In the formula, This refers to a metric for measuring data heterogeneity, used to quantify the degree of difference in data distribution between different clients. Indicates the first The number of samples in each client's local dataset.

[0035] Finally, substituting the theoretical convergence boundary, the specific formula is as follows: In the formula, It is a control parameter. This is the theoretically globally optimal model. It's the learning rate. This represents the total number of training iterations on the client's local machine.

[0036] S22, Server S The received local loss values ​​are weighted and averaged based on the data volume of each client to obtain the global loss value. The specific formula is as follows: S23. Regarding the decision-making requirements for adaptive aggregation frequency in federated learning, the server... S The server aggregates the local model parameters uploaded by each client to obtain the aggregated global model. S The convergence state of the aggregated global model is defined as the recognition framework. ,in This indicates a non-converged state, meaning that there is still room for improvement in the global model performance at the current aggregation interval, and the aggregation interval should be maintained in this case; This represents a convergence state, indicating that there is no room for improvement in the global model performance under the current aggregation interval, and the aggregation interval should be reduced in this case; S22, Server S For the global loss value sequence consisting of several consecutive rounds of global loss values ​​and the theoretical convergence boundary sequence, a sliding window of length Length is introduced for smoothing. The smoothed loss sequence and the theoretical boundary sequence are calculated using the moving average method, and the number of times two consecutive historical minimum values ​​are not reached is calculated accordingly. The specific calculation formula is as follows: In the formula, F The current sequence value, including the loss sequence and the theoretical boundary sequence, min F This is the historical minimum.

[0037] S23, Combine the two numbers As input, the two are used to activate the function. These are respectively mapped to the basic probability assignments in evidence theory. and The specific mapping formula is as follows: in, This represents the base probability after mapping.

[0038] S24. Using the Jaeger combination rule to... and Fusion yields joint evidence The specific calculation rules for the Jacobi combination rule are as follows: If , They are defined in The two mass functions on are then defined. The combined mass function is mathematically defined as follows: In the formula, the conflict coefficient , express or .

[0039] S25, Server S Based on joint evidence With fusion decision threshold Determine whether to adjust the aggregation interval, and send the adjusted aggregation interval to the client.

[0040] In S25, if If the current state is considered convergent, the aggregation interval will be adjusted. Adjusted to Otherwise, maintain the current aggregation interval. .

[0041] S3, based on aggregation interval Server S controls each client to perform local model training and aggregation, using a population optimization algorithm to update the population position vector; like Figure 3 As shown, step S3 includes: S31. Server S initializes the population size, population position vector, and velocity.

[0042] S32. Based on fitness values, the population particles are divided into three roles: explorers, followers, and vigilants. For explorer particles, a velocity update formula with random perturbation is used to update their velocity and position, enabling global exploration of the search space. For follower particles, a particle swarm optimization formula is used to update their velocity and position, ensuring overall search efficiency. For vigilant particles, simulated annealing perturbation is introduced based on their current position, and the acceptance of solutions is judged according to the Metropolis criterion. Controlled random jumps maintain population diversity and improve the global optimization capability of hyperparameter search.

[0043] Specifically, the top 20% of particles by fitness are classified as Explorers, the middle 60% as Followers, and the bottom 20% as Vigilants.

[0044] For the explorer particle, its velocity and position are updated using a velocity update formula with random perturbation, as follows: In the formula, It is a random number in the range [0,1]. For inertial weights, For individual learning factors, As a group learning factor, To explore coefficients, For the current particle in the th The speed of time, For the particle in the first The position of the era, The best position in the history of an individual particle. It is the best position in the group's history.

[0045] For follower particles, their velocity is updated using the standard particle population optimization formula, as follows: For the vigilant particle, a new solution is generated using Gaussian perturbation. x new The decision on whether to accept is based on the Metropolis criterion. x new Update temperature T( t +1)=αT( t ), where is the cooling coefficient.

[0046] S33. Calculate the fitness value of each particle and update the individual optimal position and the global optimal position.

[0047] The individuals in the population optimization algorithm include a global model composed of population position vectors. The accuracy of the global model aggregated on the server side is tested based on the current population position vectors. Calculate the fitness value f: In the formula, The test accuracy of the global model after server-side aggregation is given the current population position vector. This represents the current number of local training iterations on the client side. These are the weighting coefficients.

[0048] S34. Repeat steps S32 and S33 until the training termination condition is met, and output the optimal result.

[0049] Example 2: This embodiment provides a secure multi-party collaborative model optimization system guided by mechanistic knowledge, such as... Figure 4 As shown, it includes the following modules: The Federated Learning Environment Building Module is used to build a federated learning environment, including servers. S and N One client, one server S Initialize global model parameters and initial population position vector.

[0050] The aggregation interval calculation module is used by each client to download the initial global model parameters for federated learning, and the server... S Based on the parameters uploaded by the client and the initial population position vector, the aggregation interval of the current communication round is calculated.

[0051] The population location vector update module, based on the aggregation interval, is located on the server. S Control each client to perform local model training and aggregation, use a population optimization algorithm to update the population position vector until the training termination condition is met, output the optimized global model, and complete the collaborative model optimization.

[0052] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described in any of the above embodiments. The storage medium can be a solid-state drive, USB flash drive, server storage array, or cloud storage space, and is suitable for various federated learning platforms and intelligent system deployments.

[0053] Example 4: The dataset is the CIFAR-10 image classification dataset, which consists of 32×32 pixel three-channel color images. During data preprocessing, the pixel values ​​of CIFAR-10 are normalized to [0,1]. To simulate the non-independent and identically distributed experimental scenario in a real industrial environment, CIFAR-10 is unevenly divided according to class labels, so that different clients only have data for a subset of classes. This embodiment comprehensively evaluates the algorithm from two dimensions: model performance and communication cost. The evaluation metrics are model accuracy and the total number of aggregations required to complete training.

[0054] Experimental results are as follows Figure 5 , Figure 6 As shown, in the early stages of training, the model has a large aggregation interval, and the accuracy increases rapidly. As the number of training rounds increases, the aggregation interval decreases continuously until it reaches 1, and the accuracy continues to increase until convergence.

Claims

1. A method for optimizing a secure multi-party collaborative model guided by mechanistic knowledge, characterized in that, Includes the following steps: S1. Construct a federated learning environment, including servers. S and N One client, one server S Initialize global model parameters and initial population location vector; S2. Each client downloads the initial global model parameters for federated learning, and the server... S Based on the parameters uploaded by the client and the initial population position vector, calculate the aggregation interval of the current communication round; S3, based on aggregation interval ,server S Control each client to perform local model training and aggregation, and use a population optimization algorithm to update the population position vector; S4. Repeat steps S2 and S3 until the training termination condition is met, output the optimized global model, and complete the collaborative model optimization.

2. The optimization method for a secure multi-party collaborative model guided by mechanistic knowledge according to claim 1, characterized in that, Step S1 specifically involves: constructing a federated learning environment, including servers. S and N One client, one server S and N Each client establishes an end-to-end encrypted communication channel through the key system provided by the Key Generation Center (KGC). S Initialize global model parameters, and the server S The combination of hyperparameters to be optimized in federated learning is encoded into the initial population position vector in the population optimization algorithm. ,in For learning rate, To adjust the sliding window size, To integrate decision thresholds, This is the slope parameter of the activation function.

3. The optimization method for a secure multi-party collaborative model guided by mechanistic knowledge according to claim 2, characterized in that, In step S2, each client obtains data from the parameter server. S Download the global model parameters, load them locally for training, and upload the local model parameters and loss values; server S Based on the land model parameters and loss values ​​uploaded by the client, combined with the initial population location vector, the theoretical convergence boundary is calculated as theoretical mechanism information, and the global loss value is calculated as practical feedback information. server S The Dempster-Schaffer evidence theory is used to fuse theoretical mechanism information and practical feedback information. Based on the fusion result and the fusion decision threshold, it is determined whether to adjust the aggregation interval and output the adaptive model aggregation interval for the current communication round.

4. The optimization method for a secure multi-party collaborative model guided by mechanistic knowledge according to claim 3, characterized in that, The theoretical mechanism information is the theoretical convergence boundary of the federated learning model, and the specific calculation process includes: client... Calculate the Lipschitz constant of the local loss function based on the local loss value and local model parameters. With Smooth constant And uploaded to the server S ;server S Aggregate constants uploaded by all clients and The Lipschitz constant of the global loss function is obtained. With smoothness constant According to Lipschitz constant With smoothness constant Learning rate Total number of local training sessions and the current aggregation interval The theoretical convergence boundary is calculated based on the convergence upper bound formula of federated learning theory.

5. The optimization method for a secure multi-party collaborative model guided by mechanistic knowledge according to claim 3, characterized in that, The effective feedback information is the global model loss value. The specific calculation process includes: after the client completes local training, it calculates the loss value of its local model and uploads it to the server. S ;server S The received local loss values ​​are weighted and averaged based on the amount of data from each client to obtain the global loss value.

6. The optimization method for a secure multi-party collaborative model guided by mechanistic knowledge according to claim 3, characterized in that, The adoption of the Dempster-Schaffer evidence theory to integrate theoretical mechanism information with practical feedback information specifically includes: S21, Server S The server aggregates the local model parameters uploaded by each client to obtain the aggregated global model. S The convergence state of the aggregated global model is defined as the recognition framework. ,in This indicates a non-converged state, and the aggregation interval should be maintained in this case. This indicates a convergence state, at which point the aggregation interval should be reduced. S22, Server S For the global loss value sequence consisting of several consecutive rounds of global loss values ​​and the theoretical convergence boundary sequence, a sliding window of length Length is introduced for smoothing. The smoothed loss sequence and the theoretical boundary sequence are calculated using a moving average method. The number of times the current sequence value fails to reach the historical minimum value is calculated by comparing the current sequence value with the historical minimum sequence value. When the current sequence value is less than the historical minimum sequence value, the number of times... If it is 0, otherwise it is the previous value. Increment the value by 1; S23. Taking two exponents Φ as input, use an activation function to map the two Φs to the basic probability assignments in evidence theory. and ; S24. Using the Jaeger combination rule to... and Fusion yields joint evidence ; S25, Server S Based on joint evidence With fusion decision threshold Determine whether to adjust the aggregation interval, and send the adjusted aggregation interval to the client; like If the current state is considered convergent, the aggregation interval will be adjusted. Adjusted to Otherwise, maintain the current aggregation interval. .

7. The method for optimizing a secure multi-party collaborative model guided by mechanistic knowledge according to claim 6, characterized in that, In step S3, based on the aggregation interval ,server S Control each client to perform local model training and aggregation. After completing a preset number of federated learning training rounds, calculate the fitness value of the current population position vector and update the population position vector using a population optimization algorithm.

8. The optimization method for a secure multi-party collaborative model guided by mechanistic knowledge according to claim 7, characterized in that, The population optimization algorithm specifically includes: S31, Server S Initialize the population size, population position vector, and velocity; S32. Based on fitness values, the population particles are divided into three roles: explorers, followers, and vigilants. For explorer particles, a velocity update formula with random perturbation is used to update their velocity and position, enabling global exploration of the search space. For follower particles, a particle swarm optimization formula is used to update their velocity and position, ensuring overall search efficiency. For vigilant particles, simulated annealing perturbation is introduced based on their current position, and the acceptance of solutions is judged according to the Metropolis criterion. Controlled random jumps maintain population diversity and improve the global optimization capability of hyperparameter search. S33. Calculate the fitness value of each particle and update the individual optimal position and the global optimal position; S34. Repeat steps S32 and S33 until the training termination condition is met, and output the optimal result.

9. The optimization method for a secure multi-party collaborative model guided by mechanistic knowledge according to claim 7, characterized in that, The individuals in the population optimization algorithm include a global model composed of population position vectors. The accuracy of the global model aggregated on the server side is tested based on the current population position vectors. The fitness value f is calculated as follows: minus The value, This represents the current number of local training iterations on the client side. These are the weighting coefficients.

10. A mechanism-guided optimization system for a secure multi-party collaborative model, used to implement the secure multi-party collaborative model optimization method according to any one of claims 1 to 9, comprising the following modules: The Federated Learning Environment Building Module is used to build a federated learning environment, including servers. S and N One client, one server S Initialize global model parameters and initial population location vector; The aggregation interval calculation module is used by each client to download the initial global model parameters for federated learning, and the server... S Based on the parameters uploaded by the client and the initial population position vector, calculate the aggregation interval of the current communication round; The population location vector update module, based on the aggregation interval, is located on the server. S Control each client to perform local model training and aggregation, use a population optimization algorithm to update the population position vector until the training termination condition is met, output the optimized global model, and complete the collaborative model optimization.

Images