Federal learning layered incentive method based on personalized privacy protection

Abstract

The invention discloses a federal learning hierarchical excitation method based on personalized privacy protection. The method comprises the following steps: a task publisher TP performs conjoint analysis on privacy preference and data purchasing power of a local model owner LMO through a generative diffusion model; generating a triple incentive strategy set containing valid data volume, privacy budget and reward; each LMO selects a corresponding excitation strategy according to the type of the LMO, and worker Worker combinations meeting the effective data size requirement are screened based on excitation strategy constraint; the Worker adds noise to the original data according to the declarated privacy budget and then submits the original data to the LMO; and the LMO summarizes noise-added data to train a local model, restricts disturbance model parameters according to the privacy budget in the selected excitation strategy and then uploads the TP. According to the method, the high availability of disturbance data and a local model is ensured while the privacy of Worker local data and LMO pre-training model parameters is protected, and the balance between the overall performance of the system and the privacy protection level is guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of federated learning, and particularly relates to a federated learning hierarchical incentive method based on personalized privacy protection. BACKGROUND

[0002] The innovation of mobile computing technology and the large-scale deployment of Internet of Things devices drive the exponential growth of perception data, providing a rich multi-modal data basis for machine learning model training. However, the traditional cloud computing paradigm relies on terminal devices to upload raw data to the central server, facing two challenges: first, massive data transmission leads to network bandwidth resource overload, making it difficult to meet the low latency demand of real-time sensitive scenarios; second, data cross-domain circulation exacerbates the risk of privacy leakage, making it difficult to meet the increasingly stringent data sovereignty regulations (such as GDPR). In this context, federated learning (FL) achieves paradigm breakthrough by restructuring distributed collaboration mechanisms: local model owners (LMOs) train model parameters based on local data sets, and task publishers (TPs) aggregate distributed parameter updates using the federated averaging algorithm (FedAvg) to iteratively generate global models. Compared with traditional cloud computing, FL reduces communication overhead by 1-2 orders of magnitude through the "data local processing-model lightweight transmission" mechanism, significantly improving privacy protection strength and resource utilization efficiency.

[0003] Although federated learning has shown great potential in medical care (e.g., NVIDIA Claran FL cross-hospital joint diagnosis), Internet of Vehicles (e.g., UAV cooperative perception), and other fields, its actual deployment still faces core bottlenecks. Existing researches mostly assume that LMOs have sufficient local data to support high-quality model training. However, in reality, due to the limitations of device sensing capabilities, storage capacity, or user participation, some LMOs may face challenges such as insufficient data size or incomplete class coverage, leading to the degradation of global model convergence efficiency and generalization ability. Therefore, recruiting workers to collect additional relevant data provides an effective solution to this problem. A "TP-LMOs-Workers" three-layer collaborative architecture can be constructed: TP as the global coordinator releases tasks and designs incentive strategies; LMO as the intermediate layer participates in federated model training and recruits workers to collect key data to optimize the local model; Worker as the bottom layer executes by collecting data through sensing devices and submitting it to the corresponding LMO. This extension can alleviate the problem of data insufficiency, but how TP designs effective incentive strategies to drive LMOs to participate in FL and provide high-quality local models, and how LMOs incentivize workers to collect and upload high-value data, are still core challenges in FL hierarchical incentive research.

[0004] Currently, contract theory, auction mechanism, and game theory are the three most commonly used methods in federated learning incentive schemes. Contract theory designs contracts that meet individual rationality and incentive compatibility to guide participants to act as expected. Auction mechanism introduces a bidding process to achieve optimal resource allocation and automatic price discovery. Game theory models the strategic interaction between multiple participants to reflect the game equilibrium behavior in the incentive process. These methods are mostly limited to single-layer FL architecture and have not thoroughly considered the need to incentivize workers to collect data for LMOs. Among the existing hierarchical incentive mechanisms for federated learning, the following key problems still need further research and solution.

[0005] (1) LMO and Worker lack personalized privacy incentives

[0006] Under the framework of federated learning (FL), the core of incentive schemes is to drive participants at each layer to actively participate and contribute high-quality data or models to improve the convergence and generalization ability of the global model. However, existing incentive scheme research mainly focuses on designing incentive strategies around the compensation of data quantity and computing resources, without fully considering the privacy protection needs of LMOs when providing local model parameters. Although FL alleviates the risk of privacy leakage to some extent by avoiding the sharing of raw data, local model parameters may still be subject to inference attacks, thereby exposing data distribution or sensitive information. To this end, some research introduces differential privacy (DP) mechanism to enhance privacy protection capability by adding noise to local model parameters. However, existing incentive schemes mostly assume that all LMOs share the same privacy budget, ignoring the heterogeneity of LMOs' privacy preferences in actual scenarios. This fixed privacy budget strategy may lead to an imbalance between privacy protection level and model performance, thereby reducing the enthusiasm of LMOs to participate in FL. Under the TP-LMOs-Workers hierarchical incentive framework, the incentive problem of privacy protection becomes more complex. The addition of Workers alleviates the challenge of data insufficiency, but their data contribution behavior is significantly affected by privacy risks, and Workers may refuse to provide data due to fear of data leakage. Moreover, since different Workers have different sensitivities to privacy, if the incentive scheme cannot reasonably compensate for the corresponding privacy loss, Workers may reduce data quality (such as excessive noise) or directly exit the FL task, thereby affecting the local model training of LMOs and further weakening the performance of the global model. Therefore, how to further build an incentive scheme that takes into account the privacy preferences of different Workers on the basis of effectively motivating LMOs to participate in FL, so that TP can provide differentiated incentives according to different LMOs and LMOs can provide differentiated incentives according to different Workers' privacy needs, to optimize the balance between privacy protection and model quality, is an important research direction to improve the effectiveness of FL system incentives.

[0007] (2) Lack of consideration of hierarchical privacy budget constraints

[0008] In the federal learning hierarchical incentive scheme, the effective incentive of TP to LMO is the premise of subsequent LMO incentive Worker to contribute data. The data provided by Worker will affect the training of the local model and ultimately affect the global model. Since both Worker data and pre-trained model parameters need to be privacy disturbed, this will significantly reduce the quality of the global model. However, in the existing hierarchical architecture, TP usually only focuses on the privacy protection of the model parameters uploaded by LMO and does not directly intervene in the privacy budget setting of Worker in the LMO data procurement process, but the TP-LMO-Worker three-layer architecture has a certain coupling relationship in privacy budget control, that is, the total privacy budget of TP determines the upper limit of the privacy budget of each LMO under a certain model quality, and LMO needs to further regulate the data privacy budget of Worker under the constraint of TP privacy budget to ensure the coordinated optimization of the overall privacy protection strategy and model quality. However, existing research generally lacks systematic modeling and optimization analysis of hierarchical privacy budget constraints, resulting in obvious deficiencies in the coordinated optimization between privacy protection and model performance.

[0009] (3) Incentive challenges under information asymmetry

[0010] In the TP-LMOs hierarchical federal learning architecture, due to the significant differences in data size and privacy preferences of each LMO, it is difficult for TP to accurately evaluate their true contribution. Especially in terms of privacy protection, some LMOs may refuse to disclose the true data size or quality due to concerns about privacy leakage, or even exaggerate their data contribution to obtain more incentives. This information asymmetry behavior interferes with the reasonable evaluation of the utility of participants by TP, affecting the effectiveness of resource allocation and contribution pricing, and ultimately may lead to a decline in model aggregation performance, failing to meet the system's requirements for global model accuracy and stability. Contract theory, as a classic method to deal with information asymmetry problems, is widely used in incentive scheme design. Traditional research often uses parameter configuration methods compatible with incentives to guide participants to report their private information truthfully, thereby optimizing resource allocation and improving system efficiency. However, existing methods are mostly based on single-dimensional or low-dimensional utility functions, making it difficult to cope with the challenges brought by multi-dimensional heterogeneous factors in federal learning. On the one hand, the utility function of LMO often has high nonlinearity and coupling characteristics under multi-dimensional conditions, increasing the difficulty of theoretical derivation and solution of incentive strategy parameter design; on the other hand, traditional methods usually rely on pre-set distributions or function forms, making it difficult to adapt to dynamic and complex system environments, and easily falling into local optimum, limiting the effectiveness and global optimality of the incentive strategy. SUMMARY

[0011] To solve the above technical problems, the application provides a federated learning hierarchical incentive method based on personalized privacy protection, which constructs personalized incentive and screening methods in the two levels of TP-LMOs and LMO-Workers to optimize the balance between data quality, privacy protection and global model training efficiency.

[0012] To achieve the above object, the application provides a federated learning hierarchical incentive method based on personalized privacy protection, which includes:

[0013] The task publisher TP jointly analyzes the privacy preferences and data purchasing power of the local model owner LMO through a generative diffusion model;

[0014] A set of three-tuple incentive strategies containing effective data volume, privacy budget and reward is generated;

[0015] Each LMO selects the corresponding incentive strategy according to its own type, and screens the worker combination that meets the effective data volume requirement based on the incentive strategy constraint;

[0016] The worker submits the corresponding LMO after adding noise to the original data according to the reported privacy budget;

[0017] The LMO trains the local model by summarizing the noisy data, and uploads the TP after limiting the perturbed model parameters according to the privacy budget in the selected incentive strategy.

[0018] Optionally, the training process of the generative diffusion model includes:

[0019] Gaussian noise is gradually added to the incentive strategy vector in the forward diffusion stage;

[0020] The optimal incentive strategy distribution is recovered through a neural network in the backward diffusion stage;

[0021] The environmental state input includes the number of LMOs participating, the number of categories, the minimum privacy budget in the incentive strategy and the total incentive budget.

[0022] Optionally, the worker screening process includes:

[0023] The cost of unit effective data is calculated:

[0024] The worker combination is selected in ascending order of the value; Where the effective data volume

[0025]

[0026] Optionally, the generation conditions of the three-tuple incentive strategy include:

[0027] The privacy budget needs to be greater than the lower threshold set by the TP;​

[0028] The reward is positively correlated with the effective data amount;

[0029] The total incentive of a single LMO does not exceed a preset proportion of its data purchasing power.

[0030] Optionally, the hierarchical privacy budget constraint comprises:

[0031] The TP sets a lower limit of the model upload privacy budget for each type of LMO through a personalized incentive strategy;

[0032] The Worker voluntarily reports the data privacy budget, and the LMO requires that the total effective data amount of the data submitted after differential privacy processing be no less than the effective data amount requirement in the selected incentive strategy, thereby indirectly constraining the Worker privacy budget through the effective data amount.

[0033] Optionally, the data noise adding process comprises:

[0034] The Worker layer adds noise to the original data D i in accordance with the distribution;

[0035] The LMO layer adds Laplace noise in accordance with the differential privacy requirement.

[0036] ​The application discloses a federated learning hierarchical incentive method based on personalized privacy protection, proposes an incentive strategy design driven by a TP-LMOs layer generative diffusion model (GDM), optimizes and designs multi-dimensional incentive strategy elements by combining the generative diffusion model (GDM), generates personalized incentive strategies for different LMO characteristics on the basis of considering LMO privacy preferences and data contribution capabilities, alleviates the information asymmetry problem, and encourages LMOs to participate and contribute high-quality models. A LMO-Workers layer Worker screening scheme based on privacy preferences and data contribution is constructed, a reasonable bidding and screening method is designed by comprehensively considering the personalized privacy preferences and data contribution of Workers, Workers are encouraged to participate in data provision, and LMOs are guided to select a Worker set with higher cost performance under privacy constraints, so that the completion of the upper-layer incentive task is effectively supported. A hierarchical privacy budget constraint strategy is designed, a TP-LMOs-Workers privacy budget constraint chain is constructed: in the TP-LMOs layer, the TP sets a lower limit of the privacy budget for uploading the model for each type of LMO through the incentive strategy set, limits the differential privacy disturbance of the LMO on the local model parameters to protect privacy, and in the LMO-Workers layer, the Worker is allowed to voluntarily report the privacy budget of data collection, and the LMO requires the sum of the effective data amount of the Worker selected after differential privacy processing to be not less than the effective data amount required in the LMO selected incentive strategy, so as to indirectly constrain the matching of the privacy budget and the data contribution of the Worker through the effective data amount. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application and of the drawings illustrate the application and, together with the description, serve to explain the principles of the application. In the drawings:

[0038] Figure 1 A flowchart of a federated learning hierarchical incentive method based on personalized privacy protection is shown in the embodiment of the application.

[0039] Figure 2 A model accuracy evolution diagram with rounds is shown in the embodiment of the application. DETAILED DESCRIPTION

[0040] It should be noted that the embodiments and the features in the embodiments in the application can be combined with each other without conflict. The application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0041] It is noted that the steps shown in the flowcharts of the drawings can be performed in a computer system such as a set of computer-executable instructions, and although a logical sequence is shown in the flowcharts, in some cases the steps shown or described can be performed in an order different from that shown here.

[0042] Contract theory is the core of information economics and game theory, and is committed to solving the problem of information asymmetry in the principal-agent relationship. Its core is to design a contract that meets the incentive compatibility, so that the agent who possesses private information or takes hidden action will behave in line with the principal's goal. In modern information systems such as federated learning, edge computing, and data trading platforms, system designers often face the problem of information asymmetry when interacting with participants who have private information (such as local devices, data providers), such as not being able to obtain information about the other party's true ability, cost, data quality, and privacy preferences. As an important branch of game theory, contract theory builds a contract system that includes type identification and incentive constraints, and on the basis of protecting the individual interests of participants, it guides them to actively choose the appropriate contract and disclose information truthfully, thereby optimizing the overall efficiency of the system.

[0043] The core goal of contract design is to maximize the principal's utility while ensuring that the agent voluntarily participates and acts honestly in the presence of information asymmetry. Contract design must satisfy the following two key constraints:

[0044] Incentive compatibility (IC), each agent will choose the contract that best suits their type and will not lie about their type.

[0045] Individual rationality (IR), the utility of each agent after choosing a contract is not lower than their reservation utility (which can be set to 0).

[0046] The typical contract design process can be divided into the following three steps:

[0047] Step 1: Model definition and utility function:

[0048] First, we need to clarify the roles of the parties involved in the system and their interaction methods, and build a utility model. Assume that the system consists of a principal (such as a platform) with decision-making power and several agents (such as nodes, users) with private information. The type of each agent represents its private characteristics (such as data quality, computing power, computing ability, etc.), which is subject to a probability distribution within a certain interval. It is usually assumed to be monotonic: the higher the type, the stronger the ability and the lower the cost.

[0049] The platform designs a contract (t, r) for each type of agent, where:

[0050] t(θ): the amount of tasks (e.g., model training intensity, data uploading volume, etc.) assigned by the platform to an agent of type θ;

[0051] r(θ): the corresponding reward (e.g., money, points, or resources).

[0052] The utility function of an agent is usually defined as the reward it obtains minus the cost required to complete the task: U(θ) = r(θ) - C(θ, t(θ));

[0053] C(θ, t(θ)) represents the cost function required for an agent to complete a task, which usually satisfies the stronger the ability, the slower the task cost rises.

[0054] The goal of the platform is to maximize its utility obtained from the agent's behavior, whose utility function is V(θ) = V(t(θ)) - r(θ), which represents the value it obtains from the agent completing the task minus the reward it pays:

[0055] V(θ) = V(t(θ)) - r(θ);

[0056] where V(t(θ)) is the value brought by the task, such as training accuracy or inference coverage.

[0057] Step 2: Constraint setting:

[0058] After the model is established, in order to ensure that the agent can still "honestly" choose the contract item that matches its type in an information opaque environment, the contract design must meet two key constraint conditions:

[0059] Incentive compatibility constraint (IC):

[0060]

[0061] Ensure that each agent obtains the maximum utility from choosing the contract that matches its type, preventing agents from falsely reporting their type to seek higher income.

[0062] Individual rationality constraint (IR):

[0063] U(θ) ≥ 0;

[0064] Ensure that the utility of the agent participating is not lower than its reservation utility (usually set to 0), otherwise it will not choose to join the system.

[0065] Step 3: Optimization goal and problem solving:

[0066] After the utility model and constraints are determined, the contract design problem is transformed into a constrained optimization problem, which aims to maximize the expected revenue of the platform. The designer needs to determine the optimal task allocation function and reward function under the premise of meeting the IC and IR conditions.

[0067]

[0068] where f(θ) is the probability density function of the agent's type in the interval .

[0069] Combined with the constraint conditions in step 2, the optimal contract can be solved by mathematical methods (Lagrange multiplier method, KKT condition derivation, etc.). The final optimal contract has the property of self-selection: high-type agents will choose high-task and high-reward contract items, and the system realizes the unification of benefit maximization and agent incentive.

[0070] Generative Diffusion Models (GDMs) learn data distribution by simulating the process of gradually adding noise to data (forward diffusion) and iteratively denoising (backward diffusion), and finally realize high-quality sample generation. The core idea is to simulate the dynamic process of data from order to disorder (noise) and then to order (denoising), and learn complex data distribution.

[0071] Step 1: Forward Diffusion Process:

[0072] Forward diffusion gradually adds Gaussian noise to data through Markov chains, gradually transforming the original data into random noise.

[0073] Given the data distribution x0~q(x0), the forward process is defined as a Markov chain of T steps, and the noise injection at each step is described by the transition kernel q(x t |x t-1 ):

[0074]

[0075]

[0076] Through reparameterization, let λ t :=1-ι t , Given the input content x0, sample a Gaussian vector ε~N(0, I) to get the noise data x t at any time:

[0077]

[0078] Then x tCan be obtained by the following distribution:

[0079]

[0080] Step 2: Reverse diffusion process:

[0081] Reverse diffusion gradually restores the data distribution by training a neural network, a process that requires complex reconstruction through Markov chains. The neural network uses the knowledge it has gained to predict the noise at each step and then eliminates it. If the reverse distribution q(x t-1 |x t ) is known, one can start from x t in a normal distribution N(0, I) and then work backwards to create a new data point similar to the original dataset. However, computing q(x t-1 |x t ) requires complex calculations involving the data distribution, so an approximate estimate is made using p θ :

[0082] p θ (x t-1 |x t ) = N(x t-1 ; μ θ (x t , t), Σ θ (x t , t));

[0083] where θ are the model parameters. If the reverse formula is applied for all steps (also called trajectories), one can trace back to the original data distribution. By doing this at each step, the model can learn to predict specific features, such as the data mean and distribution at each time point. Thus, the trajectory from x T to x0 is represented as:

[0084]

[0085] After introducing conditional information g (such as the environment state) in the denoising process, p θ (x t-1 |x t , g) can be modeled as a noise prediction model, and the mean and covariance matrix of the reverse process are:

[0086] Σ θ (x t , g, t) = I t

[0087]

[0088] ​Then, the reverse diffusion chain parameterized by θ is sampled as follows:

[0089]

[0090] By minimizing the noise prediction error optimization model, the original loss function can be simplified as:

[0091]

[0092] As shown in Figure 1 The embodiment provides a federated learning hierarchical incentive method based on personalized privacy protection, which comprises the following steps:

[0093] Step 1: TP-LMOs layer: multi-dimensional incentive strategy set construction method based on generated diffusion model

[0094] Currently, there are two key problems in the federated learning incentive scheme between TP-LMOs: first, LMOs may face privacy leakage risks when uploading model parameters to TP, thereby affecting the enthusiasm of LMOs participating in federated learning; and different LMOs have different sensitivities to privacy, so if a unified incentive scheme is adopted, it may lead to unfair distribution of incentives, thereby affecting the overall cooperation efficiency. Second, LMOs may not be willing to disclose the true data size due to privacy preference differences, or some LMOs may have dishonest reporting behaviors, which makes it difficult for TP to accurately grasp the amount of effective data contributed by each LMO, resulting in a decline in the accuracy of the global model constructed by TP in the absence of true information support, and even failing to meet the performance requirements. Generally, contract theory is used to solve such information asymmetry problems, but traditional contracts are not suitable for complex network scenarios with complex utility functions, that is, multi-dimensional conditions make it difficult for traditional mathematical methods to solve contract design problems or easily fall into local optimization.

[0095] To solve the above two problems, the application adopts a scheme based on a generated diffusion model (GDM) to design personalized optimal incentive strategies, and uses GDM to solve the high dimensionality and complexity of the formulated problem. The application considers motivating LMOs to participate in federated learning by allowing each LMO to add personalized noise to its model parameters. At the same time, the application formulates differentiated incentive strategies for each LMO according to its privacy preference and the amount of data used for training, while protecting the privacy of the local model parameters of the LMO and ensuring the high availability of the local model. In addition, the generated diffusion model (GDM) is an effective tool for solving multi-dimensional optimization problems, and the incentive strategy adopted by the application needs to consider three parameter items: the privacy preference of each LMO, the amount of data used for training, and the corresponding reward. To solve the difficulty of solving in high-dimensional space by traditional contract design methods, the application proposes a multi-dimensional incentive strategy generation method based on a diffusion model. The incentive strategy designed by the TP is regarded as a multi-dimensional vector, and the diffusion-noise mechanism is used to optimize sampling in its space, so as to generate a set of incentive strategies that can maximize the utility of the TP under the privacy budget constraint, for different LMOs to select and execute. The TP explicitly stipulates the model privacy budget when generating the incentive strategy, and this budget is a hard requirement for the LMO to perform the contract. The specific process includes the following steps:

[0096] Step 1: Incentive strategy target setting (strategy triple definition and task target):

[0097] In the application, the TP wants to collect local trained model parameters from multiple local model owners (LMOs) to improve the accuracy of the global model. Due to the differences in privacy sensitivity, training ability, and resource cost of different LMOs, the TP cannot accurately perceive their true types, so it needs to design a multi-dimensional incentive strategy that meets the incentive constraints to guide each LMO to choose the most suitable strategy for itself.

[0098] Step 1.1: LMO private type modeling:

[0099] Each LMO has an unobservable private type parameter: θ m represents the data purchasing power of the LMO, and the larger the value, the more willing to purchase higher quality data, and the lower the cost in data purchase. σ n represents the privacy preference level of the LMO, and the larger the value, the more attention to privacy. These types are arranged in non-decreasing order in each dimension, and the LMO is distinguished according to these two types. Define the type of each LMO as a two-dimensional type pair:

[0100] (θ m , σ n ), m = 1, 2,..., M; n = 1, 2,..., N;

[0101] Step 1.2: Incentive strategy triple design:

[0102] TP is the LMO design incentive strategy of type (θ m ,σ n ), and the strategy triple is defined as follows:

[0103] ρ m,n =(l m,n ,∈ m,n ,r m,n );

[0104] Where: l m,n represents the effective data amount that LMO needs for local model training, which needs to be obtained by LMO from the lower layer Worker, which is equivalent to the number of samples that the data with differential privacy noise can actually contribute useful information in training; ∈ m,n represents the differential privacy budget added by LMO when uploading the model, which affects the noise intensity of the model. The larger the value, the smaller the noise, the higher the model accuracy, but the greater the risk of privacy leakage; r m,n represents the reward paid by TP to the LMO of this type, which is used to compensate for the related overheads such as purchasing data, local training, privacy loss, etc. The triple determines the training ability of LMO (through the effective data amount), the privacy leakage risk (through the privacy budget), and the incentive intensity (through the reward amount).

[0105] Step 1.3: Modeling objectives:

[0106] The objective of TP is to generate an incentive strategy that meets participation and authenticity, so that LMOs voluntarily participate (individual rationality, IR), LMOs choose strategies suitable for themselves (incentive compatibility, IC), and TP obtains the maximum utility under the budget constraint.

[0107] Step 2: Utility modeling of TP and LMO (objective function and constraint condition):

[0108] Step 2.1: Utility function of TP:

[0109] When the LMO of type (θ m ,σ n ) selects the strategy (l p,q ,∈ p,q ,r p,q ) corresponding to the LMO of type (θ p ,σ q ) and uploads the model, the utility obtained by TP is:

[0110]

[0111] Where, the model revenue is calculated as follows:

[0112]

[0113] where δ and α are the impact factors of model precision on TP's revenue based on data size determination model; γ measures the impact strength of privacy budget on model quality; log(l + α * l p,q ) indicates that increasing the amount of effective data can improve model precision, but there is diminishing marginal returns; represents the precision discount caused by uploading the local model with noise.

[0114] When the LMO of type (θ m , σ n ) selects the corresponding type strategy, the utility of TP is:

[0115]

[0116] The overall utility of TP:

[0117]

[0118] Step 2.2: Utility function of LMO:

[0119] After accepting the incentive strategy, the utility of LMO is:

[0120]

[0121] where c pur represents the unit effective data procurement cost; c com represents the calculation and communication cost; σ n * ∈ p，q represents the privacy cost. θ m represents the LMO data procurement capacity, the larger the value, the more willing to purchase higher quality data, and the lower the cost in data purchase. σ n represents the LMO privacy preference level, the larger the value, the more attention to privacy.

[0122] Step 2.3: Incentive constraint conditions

[0123] In order to make the incentive strategy run stably, it is necessary to meet:

[0124] In order to ensure that LMO obtains non-negative utility when selecting strategy (l m,n , ∈ m,n , r m,n ), it is necessary to meet the individual rationality constraint (IR):

[0125] The utility lower bound of

[0126] In order to ensure that LMO obtains non-negative utility when selecting strategy (l m,n , ∈m,n ,r m,n ) rather than other strategies, incentive compatibility (IC) needs to be satisfied:

[0127]

[0128] Step 3: Transform the incentive strategy design problem into a GDM optimization problem:

[0129] Since the traditional contract design method is difficult to solve in high-dimensional space and is easy to fall into local optimum, the incentive strategy design task is transformed into a strategy vector parameter generation problem, which is modeled and optimized by a generative diffusion model (GDM).

[0130] Step 3.1: Strategy vector space modeling:

[0131] Each incentive strategy can be regarded as a vector, and the three-tuple constitutes a strategy space:

[0132] ρ=(l m,n ,∈ m,n ,r m,m )∈R 3 ;

[0133] Step 3.2: GDM training process:

[0134] GDM adopts an iterative forward diffusion process based on the initial input data, gradually introducing Gaussian noise. Then, GDM performs a backward diffusion process through a denoising network, which iteratively approximates the true sample represented as ρ ~ Q(ρ) through a series of prediction steps, Q(ρ) representing the original data distribution. Subsequently, the denoising network is trained to reverse the noise process and restore the data and its content, thereby facilitating the generation of new data.

[0135] Step 3.2.1 Forward diffusion phase (adding noise):

[0136] In the training phase, the system first performs "perturbation" processing on the initial sample . This process gradually degrades the incentive strategy vector to an approximately pure noise distribution by gradually adding standard Gaussian noise to it. Considering ρ0 ~ Q(ρ0), the forward process in GDM can be accurately represented as a Markov process containing K steps. In the forward diffusion process, Gaussian noise is applied to the initial sample ρ0 to generate a series of samples (ρ1,..., ρ K ), then the forward diffusion process of GDM can be described as:

[0137]

[0138] where ι k∈(0,1) represents a preset diffusion intensity; k represents the number of diffusion steps; the sample of the kth step is subject to a Gaussian distribution with a mean μ k and a variance Σ k ; I is an identity matrix, indicating that each dimension has the same standard deviation.

[0139] Step 3.2.2 reverse diffusion stage:

[0140] The core of this stage is a neural network predictor for predicting the noise component contained in the current incentive strategy in each step and adjusting the strategy parameter value accordingly, that is, using a trained neural network to predict noise step by step, and recovering the incentive strategy vector ρ0from the Gaussian noise sample ρ k .

[0141] The diffusion model network is denoted as π ω (ρ|S), which maps the environment state to the incentive strategy design using weights ω. The goal of π ω (ρ|S) is to output a deterministic incentive strategy to maximize the expected cumulative reward over a series of time steps. We express the incentive strategy design as:

[0142]

[0143] where,

[0144] P ω (ρ k-1 |ρ k ,S)=N(ρ k ;μ ω (ρ k ,S,k),Σ ω (ρ k ,S,k));

[0145]

[0146] Σ ω (ρ k ,S,k)=ι k I;

[0147] In incentive strategy modeling, the environment contains various factors that affect the optimal design of the strategy, defined as S={L,M,N,∈ min ,R,(θ1...θ M ),(σ1...σ N )}, L represents the number of LMO participants in the current round, M represents the number of classifications according to the privacy preference of LMO, N represents the number of classifications according to data capability, ∈ min represents the minimum privacy budget of the incentive strategy set set by TP, and R represents the total incentive budget. In the formula, ε ωFor incentive policy generation network, λ k : = 1 - i k , Then, the inverse diffusion chain parameterized by ω is sampled as follows:

[0148]

[0149] Therefore, the task of training the incentive policy π ω in the complex high-dimensional environment S is effectively transferred to the incentive policy generation network ε ω . To train ε ω , an incentive policy quality network Q v is introduced, which maps an environment-incentive policy pair (i.e., {S, ρ}) to a value that represents the expected cumulative reward if the LMO selects an incentive policy from the current state and follows the policy. Therefore, the optimal incentive policy design is the policy that maximizes the expected cumulative utility of TP, which can be obtained by the following formula:

[0150]

[0151] The incentive policy quality network uses the traditional learning method to minimize the Bellman operator using the double Q-learning technique. Two networks, and , and three target networks, and π ω′ are constructed. Then, the network parameters v1 and v2 are optimized by minimizing the following target:

[0152]

[0153] is the discount factor, and the incentive policy design algorithm uses the denoising technique to generate the optimal policy set. Subsequently, exploration noise is introduced into the policy design and implemented to accumulate exploration experience.

[0154] Step 4: TP generates a multi-dimensional optimal incentive policy set and determines the upper limit of each local model privacy budget

[0155] After completing Step 3, the TP has constructed a generation network ε ω that can generate a multi-dimensional incentive policy set based on the current environment state by generating a diffusion model.

[0156] In this step, the TP will generate a multi-dimensional optimal incentive policy set based on the trained GDM, with the current FL task environment state S = {L, M, N, ∈ min , R, (θ1... θ M ), (σ1... σ NAs input, generate a set of optimal incentive strategy sets for different types of LMOs, and issue them for selection to each LMO.

[0157] To ensure the quality of the global model, the TP controls the privacy protection level of the entire system. When generating the incentive strategy, it sets a lower limit for the privacy budget that each LMO needs to follow when uploading the task model m，n , that is, the LMO cannot add more than the set noise to the local model.

[0158] Step two: LMO selects the appropriate incentive strategy, determines the local model privacy budget constraint and effective data volume constraint:

[0159] After completing the generation of the incentive strategy set, the TP publicly releases the optimal incentive strategy set to all LMOs. After receiving the strategy set, the LMO selects a strategy that best suits its own conditions to maximize its utility based on its private type characteristics (such as privacy preference, data capacity).

[0160] After making the incentive strategy selection, the LMO submits its response intention to the TP, and the TP confirms the binding relationship between the LMO and the selected strategy based on the received response information, officially completing the signing process. Thereafter, the LMO will complete data collection, local training, and model uploading according to the corresponding strategy requirements, and the TP will pay the corresponding incentive compensation according to the incentive strategy agreement after accepting the model.

[0161] After completing data collection and local training, the LMO must perform differential privacy processing (usually adding noise) on the uploaded local model parameters with a privacy budget of m,n , that is, it cannot add more than the constraint noise. At the same time, according to the TP incentive strategy, the LMO needs to collect training data that meets a certain effective data volume m,n . The effective data volume is related to the amount of raw data collected by each Worker and the privacy preference.

[0162] Step three: LMO-Worker layer: build a Worker screening scheme and limit the privacy disturbance level of each Worker according to the effective data volume:

[0163] In the federated learning architecture, LMO as a local model training performer needs to complete a certain amount of local training based on the incentive strategy provided by TP, however, LMO usually has limited data itself and must purchase data from Worker to complete the training task. In this process, there are the following challenges: the joining of Worker alleviates the problem of insufficient data, but its data contribution behavior is significantly affected by privacy risks, and Worker may refuse to provide data due to concerns about data leakage. In addition, due to the different sensitivity of different Workers to privacy protection, it is difficult to balance the unified incentive scheme, and if the incentive cannot reasonably compensate for the loss of privacy, the Worker may reduce the data quality by excessive noise or directly exit the federated learning task, thereby affecting the local model training effect of LMO and further weakening the performance of the global model. In order to achieve accurate matching of the heterogeneous privacy preferences and data contribution capabilities of each Worker, and balance between incentive fairness and system efficiency, it is necessary to adopt a reverse screening scheme with individualized incentive capability.

[0164] In view of the above problems, the present application designs a reverse screening incentive method based on privacy protection on the basis of the TP-LMOs incentive strategy design. Specifically, data procurement is modeled as a matching process of LMO initiating data purchase tasks, Worker voluntarily reporting participation willingness and LMO reverse screening high cost-effective Workers. Each Worker submits three parameters of data quantity, privacy budget and remuneration demand according to its own privacy preference and data collection capability. LMO calculates the effective data quantity based on the information submitted by each Worker, and then evaluates the cost-effectiveness and screens out the Worker set that meets the requirements of the TP incentive strategy. Under this scheme, LMO can dynamically balance data quality and privacy protection, achieve data procurement for local training at a lower cost, and support the performance and privacy target collaborative optimization of the overall FL system. Taking an LMO of any type (θ m ,σ n ) as an example, the specific process includes the following steps:

[0165] Step 1: Effective data quantity calculation method:

[0166] The effective data quantity is usually related to the original data quantity and the noise variance, and the noise variance is related to the privacy budget. In differential privacy, the noise variance is inversely proportional to the square of the privacy budget. Therefore, after each Worker perturbs the original data and uploads it, the effective data quantity is:

[0167]

[0168] c represents the noise coefficient related to the task (constant, reflecting the sensitivity of the task and the differential privacy mechanism used; in many literature, the value is 1 or other empirical constant); di denotes the amount of raw data provided by Worker i; ∈ i denotes the privacy budget of Worker i.

[0169] Step 2: Utility function design:

[0170] When the type is (θ m , σ n ), the LMO accepts the incentive policy corresponding to its type, and its task objective in this round is to train the model using at least l m,n pieces of valid data, and upload the model to the TP after adding noise to the model with a privacy budget higher than ∈ m,n . To achieve this goal, the LMO will select from the set of Workers who apply to sell data. The selection strategy aims to maximize its own utility and minimize data procurement costs while meeting the constraints of the corresponding incentive policy and encouraging Workers to participate.

[0171] Workers submit bids based on their own circumstances, including the amount of data d i they are willing to provide and the acceptable privacy budget ∈ i , and expect to obtain payment p i from the LMO. The utility function of the Worker is:

[0172]

[0173] where, denotes the unit data collection cost of Worker i; denotes the unit privacy leakage cost of Worker i, and the participation constraint of Worker i is the lower bound of the utility of i, which is generally taken as 0.

[0174] After the LMO selects the incentive policy corresponding to the type, the utility is:

[0175]

[0176] where, the unit valid data procurement cost c pur comes from the average cost composed of the bid prices of the winning Workers.

[0177] The specific formula is:

[0178] where, W m，n denotes the set of winning Workers of the LMO of type (θ m , σ n ); denotes the total bid price of the winning Workers.

[0179] Step 3: LMO initiates a screening task:

[0180] Type of (θ m , σ n ) LMO after receiving the incentive policy (l m，n , ∈ m，n , r m，n ) formulated by TP, carries out the local data procurement process. The procurement data needs to meet the total effective data volume not less than l m,n , but does not force the unified privacy budget of the Worker, allowing the Worker to autonomously decide the privacy budget according to its own privacy preference. The LMO broadcasts the screening task to the Worker set under its jurisdiction, and explicitly requires each Worker to submit a bid triplet: b i = (d i , ∈ i , p i ).

[0181] Step 4: Worker submits participation application:

[0182] Each Worker submits an application triplet: b i = (d i , ∈ i , p i ) to the LMO based on its own resource situation and privacy sensitivity. The Worker aims to maximize the probability of winning the bid while ensuring its own utility is non-negative.

[0183] Step 5: LMO calculates the cost-effectiveness and selects the winning bidder:

[0184] When screening the Worker, the LMO cannot simply look at the original data volume d i , but must follow the effective data volume i under the privacy budget ∈ . The effective data volume formula can reflect the constraints on the Worker's privacy budget and original data volume. For each Worker i, according to its application information, calculate its effective data volume:

[0185]

[0186] Then, evaluate the unit effective data procurement cost of each Worker i:

[0187]

[0188] Where, The smaller the , the smaller the cost consumed per unit of effective data.

[0189] Sort the values ​​from small to large (the lower the cost per unit of effective data volume, the better); select Workers in turn to form a set W m，n , and must meet If there are multiple combinations that meet the conditions, the one with the lowest total reward will be selected.

[0190] Step 4: Workers perform personalized perturbations on the data based on the application information and upload it to LMO:

[0191] Worker i applies privacy budget ∈ i For data volume d i The data is processed with differential privacy and noise:

[0192]

[0193] in After adding noise, the data is uploaded to LMO, and Worker i receives reward p i .

[0194] Step 5: LMO completes local model training based on the perturbed data and uploads the perturbed model

[0195] LMO merges all the winning Worker data to construct a training set:

[0196]

[0197] Use this dataset to train the local model and follow the ∈ m,n After adding noise to the model parameters, upload it to TP to complete this round of tasks.

[0198] TP receives the noise model parameters uploaded by each LMO, evaluates and aggregates them, and obtains the global model.

[0199] This experiment is based on the federated learning hierarchical incentive framework (TP-LMO-Worker) proposed for user privacy protection, and uses real data sets (EMNIST and Fashion-MNIST) for numerical simulation to evaluate the impact of four methods on system performance. The system architecture includes 1 task publisher (TP), 5 local model organizers (LMO) and 100 terminal participants (Worker), and each LMO manages 20 Workers. The experiment compares the performance of four schemes, namely uniform pricing, discriminatory pricing, hierarchical incentive mechanism (HUMA) and the method of the present invention, in terms of model accuracy. The simulation uses a CNN model for federated communication. The training parameters include a learning rate of 0.01, a batch size of 32, and non-IID data partitioning (Dirichlet distribution α = 0.5). The experiment focuses on simulating the curve of the accuracy change with the number of training rounds under different strategies, such as Figure 2 shown.

[0200] Uniform Pricing: TP provides the same unit data price to all LMOs, and LMOs also purchase Worker data at a uniform price standard without considering the personalized characteristics (such as data capacity or privacy preference) of LMOs or Workers for differential pricing.

[0201] Discriminative Pricing: TP sets different prices for each unit of data for different types of LMOs. LMOs set different prices for each unit of data for different types of Workers. In actual scenarios, TP cannot accurately know the private attributes of each LMO, and LMOs cannot accurately know the private attributes of each Worker, such as data collection capacity, data quality, etc., making it difficult to implement this pricing method.

[0202] Hierarchical Incentive Pricing (HUMA): HUMA pricing is different from traditional methods. TP determines the total payment τ to LMOs by means of Stackelberg game. LMOs use contract theory to develop differential rewards and data volume requirements for different Workers based on their own circumstances and the game results with TP, in accordance with individual rationality and incentive compatibility constraints, to achieve pricing of Worker data.

[0203] An application example of the present application is as follows:

[0204] 1. Scene background

[0205] In a provincial smart medical system, a regional medical data platform hopes to jointly train a lung disease recognition model based on X-ray images with several third-level hospitals and community health service centers in the province. The platform does not directly access the original data, but coordinates the third-level hospitals to participate in training through community medical sites in the jurisdiction to ensure model effectiveness and protect patient privacy.

[0206] Each community medical site stores local patients' lung X-ray images, which are highly privacy-sensitive and differ in device performance, data volume, privacy protection willingness, etc. If each site is forced to upload perturbed models or anonymous data, it is easy to cause a decrease in participation willingness and even cause training failure. To solve this problem, the platform introduces a personalized privacy protection method and an incentive hierarchical screening strategy to improve training efficiency and privacy security level.

[0207] TP (Task Publisher): Regional medical platform, sets task goals and reward rules, and aggregates global models;

[0208] LMO (Local Model Organization): A tertiary hospital, organizes community clinics under its jurisdiction to participate in training;

[0209] Worker (Edge Terminal): Community medical sites or mobile diagnostic devices, holds raw image data.

[0210] 2. Implementation Case

[0211] Step 1: The platform formulates various levels of participation programs and distributes them to tertiary hospitals

[0212] The regional medical platform sets the model accuracy target for this task and formulates various levels of participation programs through the generation of diffusion models. For example, the first level program requires uploading about 10,000 perturbed images, corresponding to a minimum upload model privacy budget of 1.2, and the upper limit of compensation is 20,000 yuan; the second level program requires uploading about 8,000 perturbed images, corresponding to a minimum upload model privacy budget of 1.0, and the upper limit of compensation is 17,000 yuan; the third level program requires uploading about 6,000 perturbed images, corresponding to a minimum upload model privacy budget of 0.8, and the upper limit of compensation is 12,000 yuan. These programs clearly specify the amount of valid training data each tertiary hospital needs to submit, and stipulate that a certain degree of differential privacy perturbation must be added before uploading the local model to protect the data source. Each program corresponds to different reward incentive standards, which are uniformly announced by the platform. Tertiary hospitals can choose the most suitable program according to their own resource status and the ability of subordinate medical sites, and submit confirmation to the platform. For example: a certain people's hospital in Chengdu has strong data capacity and relatively loose privacy control, so it is automatically recommended to participate in the first level program; while a certain people's hospital in Liangshan Prefecture has high privacy preference, so it is recommended to participate in the third level program with low intervention privacy budget.

[0213] Step 2: Tertiary hospitals initiate task notifications to medical sites in their jurisdiction

[0214] After completing the program selection, the tertiary hospital sends a task announcement to the community medical sites under its jurisdiction. The announcement lists the requirements for participating in this training, including the required data volume range, acceptable privacy protection level (i.e. model perturbation degree), and compensation incentives that each site can apply for. Community medical sites decide whether to participate based on their own image quantity, data sensitivity, and computing capacity. Then fill in the data provision willingness, the acceptable perturbation degree, and the desired compensation, etc. For example: a certain community health service center in Jinjiang District of Chengdu expresses its willingness to contribute 400 images, upload perturbed models with privacy budget ∈ = 1.2, and offer a price of 3 yuan per image.

[0215] Step 3: Tertiary hospitals screen medical sites to form training alliances

[0216] After receiving the response information of the sites, the tertiary hospital evaluates the filling content of all sites. The sites with higher data quality, willingness to provide training results under controllable privacy budget, and reasonable price are selected to form the training dataset. This process ensures that sufficient training data can be obtained without violating the upper limit control requirements of the platform in terms of privacy. The selected sites will receive qualification confirmation and data upload channel. For example: According to the price of all sites, a hospital in the high-tech zone selected 4 sites with large data volume and reasonable privacy budget, and after combination, about 10,000 images required by the first level scheme can be met.

[0217] Step 4: Site completes perturbation and submits perturbed data

[0218] The selected sites will receive qualification confirmation and data upload channel. After obtaining the confirmation, each community medical site perturbs the local data according to the privacy budget in the application, and then uploads the perturbed data to the tertiary hospital.

[0219] Step 5: Tertiary hospital completes local model training and perturbation and uploads to platform

[0220] The tertiary hospital aggregates the perturbed data submitted by all sites, and uses the perturbed X-ray images to train the model. After training, the tertiary hospital perturbs the model parameters or gradients according to the privacy budget promised by itself, and then submits the perturbed model results to the platform. The whole process does not need to upload the original images, which greatly reduces the risk of patient privacy leakage.

[0221] Step 6: Platform completes perturbed model aggregation

[0222] After receiving all the perturbed models submitted by the hospitals, the platform performs aggregation processing, and integrates the global model uniformly. Then the global model is issued for iterative training. When the training accuracy reaches the set standard, the platform will pay the corresponding rewards to the hospitals according to the incentive amount in the task scheme. The tertiary hospital can further pay the compensation to the sites according to the unit price in the site application.

[0223] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed in the present application can be easily thought by those skilled in the art, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A hierarchical incentive method for federated learning based on personalized privacy protection, characterized by: include: The task publisher TP conducts a joint analysis of the privacy preferences and data purchasing power of the local model owner LMO through a generative diffusion model; Generate a set of incentive strategies consisting of a triplet of effective data volume, privacy budget, and reward; Each LMO selects an incentive strategy based on its own type and selects a worker combination that meets the effective data volume requirements based on the incentive strategy constraints. The Worker adds noise to the original data according to the declared privacy budget and submits the corresponding LMO; LMO aggregates the noisy data to train the local model, perturbs the model parameters according to the privacy budget limit in the selected incentive strategy, and then uploads TP.

2. The hierarchical incentive method for federated learning based on personalized privacy protection according to claim 1, characterized in that: The training process of the generative diffusion model includes: In the forward diffusion stage, Gaussian noise is gradually added to the incentive strategy vector; The back-diffusion phase recovers the optimal incentive strategy distribution through the neural network; The environment state input includes the number of LMO participants, the number of categories, the minimum privacy budget in the incentive strategy, and the total incentive budget.

3. The hierarchical incentive method for federated learning based on personalized privacy protection according to claim 1, characterized in that: The Worker screening process includes: Calculate the unit effective data cost: according to Select the Worker combination in ascending order of values; The effective data volume 4. The hierarchical incentive method for federated learning based on personalized privacy protection according to claim 1, characterized in that: The generation conditions of the triplet incentive strategy include: The privacy budget must be greater than the lower threshold set by TP; The compensation is positively correlated with the amount of effective data; The total amount of incentives for a single LMO does not exceed a preset proportion of its data purchasing power.

5. The hierarchical incentive method for federated learning based on personalized privacy protection according to claim 1, characterized in that: The hierarchical privacy budget constraints include: TP sets the privacy budget lower limit for model upload for each type of LMO through personalized incentive strategies; Workers independently declare their data privacy budget. During LMO screening, the total amount of valid data submitted after differential privacy processing must be no less than the valid data requirement in the incentive strategy selected by LMO, indirectly constraining the Worker privacy budget through the amount of valid data.

6. The hierarchical incentive method for federated learning based on personalized privacy protection according to claim 1, characterized in that: The data noise adding process includes: The Worker layer processes the original data D i Add a distributed noise; The LMO layer adds Laplace noise to the model parameters that meets the requirements of differential privacy.

Images