A method for estimating the contribution of participant models in vertical federated learning

Abstract

The present invention discloses a method for estimating the contribution of a participant model in vertical federated learning, and belongs to the technical field of cooperation and coordination between intelligent agents. The method estimates the contribution of participants to the model performance before the start of vertical federated learning training based on the contribution of participants to mutual information. It calculates the mutual information between features and labels distributed among different participants by training a distributed neural network in the form of vertical federated learning to avoid introducing additional privacy risks beyond the vertical federation itself. At the same time, a scalar attention mechanism is integrated in the network, and the attention weight is regarded as the contribution of the participant, so as to predict the contribution of all participants in one network training, and evaluate the redundancy of the contribution to achieve fair contribution evaluation.

Description

Technical Field

[0001] The present invention belongs to the technical field of cooperation and coordination between intelligent agents, and specifically relates to the cooperation and coordination of participants in vertical federated learning, and in particular to a method for estimating the contribution of participant models in vertical federated learning. Background Art

[0002] Vertical federated learning is widely used in many fields, improving the intelligence and accuracy of risk control, disease diagnosis, advertising, insurance pricing, production optimization, inventory management, and personalized learning. Through vertical federated learning, the task publisher and the participants work together to train the machine learning model without leaking their original data. The task publisher uses the vertical federated learning model to provide services to customers to increase business profits. In order to motivate participants to actively join vertical federated learning, the task publisher will distribute corresponding rewards according to the proportion of participants' contribution to model performance. Existing methods for contribution estimation can identify important features or participants during training and provide evaluation results at the end of training. However, this delay in evaluation infringes on the rights of participants. Although some participants may not be satisfied with the rewards, they can only accept this result because the training has been completed. Therefore, before the start of vertical federated learning, the reward distribution should be determined in advance based on the estimated contribution ratio, which is the contribution estimation, so that participants can decide whether to join based on their satisfaction with the promised rewards. However, there are the following problems to be solved in contribution estimation:

[0003] 1) There is a conflict between security and data access. The mutual information between labels and features is positively correlated with the final model performance, so the contribution of participants can be estimated by their contribution to the mutual information. However, the estimation of mutual information requires access to the original data, which contradicts the original intention of vertical federated learning. Although it is feasible to protect the transmitted features or labels by means such as multi-party computing and homomorphic encryption, it will incur high communication costs and may fail under certain conditions, thereby introducing additional privacy risks in addition to the inherent risks of vertical federated learning (i.e., gradient transmission). Therefore, contribution estimation cannot introduce privacy risks brought by additional encryption means, and it is necessary to ensure the security of the data.

[0004] 2) There is a conflict between the accuracy and efficiency of contribution estimation. It is inaccurate to simply equate the contribution of a participant with the mutual information between its features and the label, because the features of different participants may be correlated or redundant. In order to solve the problem of redundant contributions, the Shapley value in game theory can be introduced. The Shapley value of a participant is its average marginal contribution in all addition sequences. Although the Shapley value ensures fairness and accuracy, its computational cost is huge and requires 2 calculations for each participant. n-1The calculation of the submarginal contribution, where n is the total number of participants. Sampling the order in which participants are added instead of enumerating all cases can reduce the calculation time, but this method sacrifices the accuracy of the contribution measurement and still requires repeated calculations. Therefore, the contribution estimate needs to be close to the Shapley value while avoiding repeated calculations of the contribution of each participant. Summary of the invention

[0005] Purpose of the invention: In view of the security, accuracy and efficiency problems existing in the above-mentioned prior art of vertical federated contribution estimation, the present invention provides a method for estimating the contribution of participant models in vertical federated learning, thereby realizing an accurate, efficient and secure contribution estimation method.

[0006] Technical solution: An efficient and secure contribution estimation method in vertical federated learning, which estimates the contribution of participants to model performance based on their contribution to mutual information. This method trains a distributed neural network in the form of vertical federated learning and calculates the mutual information between features and labels distributed among different participants to avoid introducing additional privacy risks beyond the vertical federation itself. At the same time, a scalar attention mechanism is integrated into the network, and the attention weight is regarded as the contribution of the participant, so as to predict the contribution of all participants in one network training and evaluate the redundancy of the contribution.

[0007] The method comprises the following steps:

[0008] (1) Determine the participants who will contribute to the estimation, align their samples, and then jointly build a distributed network in the form of vertical federated learning and initialize the network parameters;

[0009] Among them, each participant constitutes a set N = {P i} i∈N , the task publisher P n Also included in N is a special participant with label y. The sample alignment dataset of participants in N is M, and the mth data in M ​​is (x m ,y m ),in

[0010] (2) Extract a sample batch B from M and publish all sample IDs contained in batch B to all participants;

[0011] (3) Participants i} i∈N Its local characteristics As input, the local network outputs a vector representation of the feature and transmit it to P n , The calculation formula is as follows:

[0012]

[0013] where h i For participant P i local network;

[0014] (4) The scalar attention module in the aggregation network is used as The encoder obtains its encoding result {z m} m∈B , and its calculation formula is as follows:

[0015] z m =h a ({h i} i∈N θ a ),

[0016] where h a is the network of scalar attention modules, θ a are the parameters of the network;

[0017] (5) m} m∈B and {y m} m∈B As input, the network output function R is estimated through mutual information θ (x m ,y m ), and calculate the expectation R θ (x m ,y m ) is calculated as follows:

[0018] R θ (x m ,y m )=h f (z m ,y m θ a ),

[0019] where h f is the mutual information estimation network, θ f are the parameters of the network;

[0020] (6) Disrupt B to obtain {z m} m∈B and Estimating network output function through mutual information And calculate the expectation The calculation formula is as follows:

[0021]

[0022] (7) Calculate the lower bound of the mutual information estimated for the batch of data B based on the expectations in (5) and (6): will maximize As the goal of distributed networks, I Θ The calculation formula is as follows:

[0023]

[0024] Among them, I Θ is the set of all participant characteristics ∪ i∈N x i The mutual information with label y, Θ is the parameter of the distributed network, Θ=({ω i} i∈N ,θ a ,θ f ).

[0025] (8) Update the parameters θ of the aggregation network a and θ f ;

[0026] (9)P n calculate and transmit it to each participant, The calculation formula is as follows:

[0027]

[0028] (10) Each participant shall Compute the gradient of the local network And update the network parameters, The calculation formula is as follows:

[0029]

[0030] (11) Repeat steps (1) to (10) until the network converges. That is ∪ i∈N x i The mutual information with the label y, while saving the weights of the scalar-level attention module as the contribution of each participant.

[0031] Furthermore, the distributed network of vertical federated learning constructed in step (1) consists of two parts: the local network of each participant {h i} i∈N , the task publisher P n The aggregate network h=(h a ,h f ), where h a is a scalar attention network, h f Estimation network for mutual information.

[0032] Furthermore, the scalar attention module in step (4) can learn a scalar in the network to represent the contribution. Its attention score The transposed vector b is added to the calculation T , to ensure that it is a scalar. At the same time, the label y m is used as the key of the attention module and is passed To calculate the attention score. Among them, σ is the activation function, A1 is the m The matrix mapped to the latent space, A2 is the matrix that Matrix that maps to latent space. The calculation formula is as follows:

[0033]

[0034] Furthermore, the scalar attention module in step (4) calculates the i A set of attention weights P i Contribution i for The mean of the fitted Gaussian distribution. In order to ensure the fairness of contribution estimation, the task publisher P n The attention weights of need to be removed, so P i The real contribution for The normalized result.

[0035] Furthermore, the function R in steps (5) and (6) θ is the function R in the dual f-divergence representation of the mutual information lower bound. The calculation formula of the dual f-divergence of the mutual information lower bound is as follows:

[0036]

[0037] The above formula indicates that we take the appropriate function R to maximize the following formula. Among them, I(Y;X) represents the mutual information between random variables Y and X. is a family of functions in R P XY represents the joint distribution between X and Y, represents the product of the marginal probability of X and the marginal probability of Y.

[0038] Beneficial effects: Compared with the prior art, the substantial progress and significant effects of the present invention are as follows:

[0039] 1) The method of the present invention is based on the training of distributed neural networks in the form of vertical federated learning, calculates the mutual information between features and labels distributed among different participants, and measures the contributions of the participants to avoid introducing additional privacy risks beyond the vertical federation itself and ensure the security of the data;

[0040] 2) The method of the present invention integrates a scalar-level attention mechanism in the mutual information calculation network and regards the attention weight as the contribution of the participant, so that the contribution of all participants can be predicted in one network training and the redundancy of the contribution can be evaluated, thereby improving the efficiency and accuracy of contribution estimation;

[0041] 3) The method described in the present invention complies with the properties of Shapley value such as balance, zero element and symmetry, and can fairly evaluate contributions, ensuring that the contribution of each participant is fairly recognized and rewarded. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the structure of the edge server group constructed by the method of the present invention;

[0043] Figure 2 The figure shows a service topology diagram of a regional edge server computing network in an embodiment;

[0044] Figure 3 is a load redistribution ratio diagram of edge server k in the embodiment. DETAILED DESCRIPTION

[0045] In order to explain the technical solution disclosed in the present invention in detail, further elaboration is made below in conjunction with the accompanying drawings and specific embodiments of the specification.

[0046] In view of the contradiction between data access and privacy protection in contribution estimation in vertical federated learning, as well as the contradiction between the accuracy and efficiency of contribution estimation, which leads to the problem that contributions cannot be estimated safely and efficiently, the present invention provides an accurate, efficient and secure contribution estimation method. This method estimates the contribution of the participants to the model performance based on their contribution to the mutual information. By training a distributed neural network in the form of vertical federated learning, the mutual information between features and labels distributed among different participants is calculated to avoid introducing additional privacy risks. At the same time, a scalar attention mechanism is integrated in the network, and the attention weight is regarded as the contribution of the participant, so as to predict the contribution of all participants in one network training and evaluate the redundancy of the contribution.

[0047] In the method described in the present invention, the method estimates the contribution of participants based on a distributed neural network in the form of vertical federated learning. Different from the existing work of estimating the contribution of participants, this method estimates the contribution directly by training the vertical federated network, avoiding additional privacy leakage except for gradient transmission, and the scalar attention module in the network can simultaneously predict the contribution of all participants and evaluate the redundancy of the contribution, thus achieving efficient contribution evaluation. Based on the above technical solution content and the existing technology that should be familiar to those skilled in the art, further, combined with Figure 1 The implementation process of the method described in the present invention is described in detail below.

[0048] Step 1: Build a distributed network in the form of vertical federated learning.

[0049] After identifying each participant involved in the contribution estimation and aligning their samples, the participants jointly build a distributed network in the form of vertical federated learning and initialize the network parameters.

[0050] Among them, each participant constitutes a set N = {P i} i∈N , the task publisher P n Also included in N is a special participant with label y. The sample alignment dataset of participants in N is M, and the mth data in M ​​is (x m ,y m ),in

[0051] The distributed network of this vertical federated learning form consists of two parts: the local network of each participant {h i} i∈N , the task publisher P n The aggregate network h=(h a ,h f ), where h a is a scalar attention network, h f Estimation network for mutual information.

[0052] Step 2: Extract sample batches and publish sample IDs, preparing for forward and backward iterations of the network.

[0053] Draw a sample batch B from M and publish all sample IDs contained in batch B to all participants;

[0054] Step 3: The participants calculate the vector representation of local features locally.

[0055] Participants i} i∈N Its local characteristics As input, the local network outputs a vector representation of the feature and transmit it to P n , The calculation formula is as follows:

[0056]

[0057] where h i For participant P i local network;

[0058] Step 4: Encode the participant’s feature vector via a scalar-level attention module.

[0059] The scalar attention module in the aggregation network is used as The encoder obtains its encoding result {z m} m∈B , and its calculation formula is as follows:

[0060] z m =h a ({h i} i∈N θ a ),

[0061] Among them, h a is the network of scalar attention modules, θ a are the parameters of the network.

[0062] In the formula, z m The specific calculation formula is as follows:

[0063]

[0064] To ensure the effectiveness of the attention mechanism, it is necessary to ensure so Actually, it is the attention score The normalized result is The calculation formula is as follows:

[0065]

[0066] Furthermore, it is necessary to design a suitable scalar-level attention module so that it can learn a scalar in the network to represent the contribution. The transposed vector b is added to the calculation T , to ensure that it is a scalar. At the same time, the label y m is used as the key of the attention module and is passed To calculate the attention score. Among them, σ is the activation function, A1 is the m The matrix mapped to the latent space, A2 is the matrix that Matrix that maps to latent space. The calculation formula is as follows:

[0067]

[0068] Furthermore, the scalar attention module calculates the i A set of attention weights P i Contribution i for The mean of the fitted Gaussian distribution. In order to ensure the fairness of contribution estimation, the task publisher P n The attention weights of need to be removed, so P i The real contribution for The normalized result.

[0069] Step 5: Fit the function R in the mutual information lower bound θ (x m ,y m ).

[0070] {z m} m∈B and {y m} m∈B As input, the network output function R is estimated through mutual information θ (x m ,y m ), and calculate the expectation R θ (x m ,y m ) is calculated as follows:

[0071] R θ (x m ,y m )=h f (z m ,y m θ a ),

[0072] where h f is the mutual information estimation network, θ f are the parameters of the network.

[0073] Step 6: Fitting a function in the mutual information lower bound

[0074] Disrupt B to get {z m} m∈B and Estimating network output function through mutual information And calculate the expectation The calculation formula is as follows:

[0075]

[0076] In steps 5 and 6, function R θ is the function R in the dual f-divergence representation of the mutual information lower bound. The calculation formula of the dual f-divergence of the mutual information lower bound is as follows:

[0077]

[0078] Where I(Y;X) represents the mutual information between random variables Y and X. is a family of functions in R P XY represents the joint distribution between X and Y, represents the product of the marginal probability of X and the marginal probability of Y.

[0079] In the formula, the mutual information network is used to calculate Fitting Calculated using mutual information computing network Fitting

[0080] Step 7: Calculate the mutual information lower bound

[0081] Calculate the lower bound of the mutual information estimated for this batch of data B based on the expectations in steps 5 and 6 will maximize As the goal of distributed networks, I Θ The calculation formula is as follows:

[0082]

[0083] Among them, I Θ is the set of all participant characteristics ∪ i∈N x i The mutual information with label y, Θ is the parameter of the distributed network, Θ=({ω i} i∈N ,θ a ,θ f ).

[0084] Step 8: Update the parameters θ of the aggregation network a and θ f .

[0085] Step 9: Calculate and transmit each participant’s

[0086] P n calculate and transmit it to each participant, The calculation formula is as follows:

[0087]

[0088] Step 10: Update the participant's local network parameters.

[0089] Each participant Compute the gradient of the local network And update the network parameters, The calculation formula is as follows:

[0090]

[0091] Step 11: Repeat the forward and backward iterations of the network.

[0092] Repeat steps 1-10 until the network converges. That is ∪ i∈N x i The mutual information with the label y, while saving the weights of the scalar-level attention module as the contribution of each participant.

[0093] Combination Figure 2 As shown, the performance of the mutual information estimation is tested and the performance comparison of different mutual information estimation methods is shown. Since the accurate calculation of mutual information values ​​is only applicable to specific distributions and discrete data, the present invention randomly generates 6000 data points as a data set based on Gaussian distribution. These data points are generated using 20 different correlation coefficients, which are uniformly sampled from the interval [-1,1]. Figure 3 From the different curves in FIG. 1 , it can be seen that compared with the traditional mutual information estimation method, the mutual information estimation method of the present invention is obviously closer to the true mutual information value.

[0094] Combination Figure 3 As shown in Figure , the performance of the neural network-based contribution estimation was tested, using the true Shapley value of the mutual information as the benchmark for the method. Figure 3 From the scatter points in , there is an obvious positive correlation between the contribution of the mutual information of the participants' features estimated by the present invention and their true Shapley value, which confirms the effectiveness of the method.

[0095] The redundant contribution evaluation performance was tested as shown in Table 1. Six public data sets, including letter recognition, obesity index, character recognition, dry bean types, and red wine quality index, were selected, and two participants P4 and P4' with completely redundant features were set. From the data in the last two columns of the table, the participants P4 and P4' estimated by the method are exactly the same, which confirms the effectiveness of the redundant contribution evaluation of the method and illustrates the fairness of the contribution estimation of the method.

[0096] Table 1. Performance of participant redundant contribution evaluation

[0097]

Claims

1. A method for estimating the contribution of participant models in vertical federated learning, characterized by: The method includes training a neural network for computing the mutual information between features and labels between different participants in a longitudinal federated learning process, integrating a scalar attention mechanism in the neural network, and considering the attention weight as the contribution of the participants; The method comprises the following steps: (1) Determine the participants who will contribute to the estimation, align their samples, and then jointly build a distributed network in the form of vertical federated learning and initialize the network parameters; Each participant constitutes a set N = {P i } i∈N , the task publisher P n It is also included in the set N and is defined as a special participant with label y. After the samples corresponding to the participants in the set N are aligned, the dataset M is obtained. The mth data in the dataset M is represented as (x m ,y m ),in (2) Extract a sample batch B from the dataset M and publish all sample IDs contained in the batch B to all participants; (3) Participants {P i } i∈N Its local characteristics As input, the local network outputs a vector representation of the feature And transmit it to the task publisher P n , The calculation formula is as follows: where h i For participant P i The local network, ω i Parameters representing the participant's local network; (4) The scalar attention module in the aggregation network is used as The encoder obtains its encoding result {z m } m∈B , and its calculation formula is as follows: z m =h a ({h i } i∈N ;θ a ), Among them, h a is the network of scalar attention modules, θ a are the parameters of the network; In the formula, z m The specific calculation formula is as follows: To ensure the effectiveness of the attention mechanism, it is necessary to ensure is the attention score The normalized result is The calculation formula is as follows: (5) m } m∈B and {y m } m∈B As input, the network output function R is estimated through mutual information θ (x m ,y m ), and calculate the expectation R θ (x m ,y m ) is calculated as follows: R θ (x m ,y m )=h f (z m ,y m ;θ a ), where h f is the mutual information estimation network, θ f are the parameters of the network; (6) Shuffle sample batch B to obtain {z m } m∈B and Estimating network output function through mutual information And calculate the expectation The calculation formula is as follows: (7) Calculate the lower bound of the mutual information estimated for the batch of data B based on the expectations in (5) and (6): will maximize As the goal of distributed networks, I Θ The calculation formula is as follows: Among them, I Θ is the set of all participant characteristics ∪ i∈N x i The mutual information with label y, Θ is the parameter of the distributed network, Θ=({ω i } i∈N ,θ a ,θ f ); (8) Update the parameters θ of the aggregation network a and θ f ; (9)P n calculate and transmit it to each participant, The calculation formula is as follows: (10) Each participant shall Compute the gradient of the local network And update the network parameters, The calculation formula is as follows: (11) Repeat steps (1) to (10) until the network converges. That is ∪ i∈N x i The mutual information with the label y, while saving the weights of the scalar-level attention module as the contribution of each participant.

2. The method for estimating the contribution of participant models in vertical federated learning according to claim 1, characterized in that: The distributed network of vertical federated learning constructed in step (1) includes each participant’s local network {h i } i∈N and the task publisher P n The aggregate network h=(h a ,h f ), where h a is a scalar attention network, h f Estimation network for mutual information.

3. The method for estimating the contribution of participant models in vertical federated learning according to claim 1, characterized in that: The scalar-level attention module described in step (4) is learned through the network to obtain a scalar to represent the contribution. The attention score Add the transposed vector b to the calculation T , to ensure that it is a scalar; at the same time, the label y m is used as the key of the attention module and is passed To calculate the attention score, where σ is the activation function and A1 is the m The matrix mapped to the latent space, A2 is the matrix that Matrix mapping to latent space; Attention score The calculation formula is as follows:

4. The method for estimating the contribution of participant models in vertical federated learning according to claim 1 or 3, characterized in that: The scalar attention module in step (4) involves calculating for each participant P i A set of attention weights P i Contribution i for The mean of the fitted Gaussian distribution; In order to ensure the fairness of contribution estimation, the task publisher P n The attention weights of need to be removed, so P i The real contribution for The normalized result.

5. The method for estimating the contribution of participant models in vertical federated learning according to claim 1, characterized in that: The function R in steps (5) and (6) θ is the function R in the dual f-divergence representation of the mutual information lower bound. The calculation formula of the dual f-divergence of the mutual information lower bound is as follows: Where I(Y;X) represents the mutual information between random variables Y and X. is a family of functions in R P XY represents the joint distribution between X and Y, represents the product of the marginal probability of X and the marginal probability of Y.

Images