Federal learning fair cooperation system and method based on multi-dimensional dynamic evaluation and adaptive customer selection

Abstract

The invention belongs to the technical field of federated learning, relates to a multidimensional dynamic aggregation federated learning framework, and aims to solve the problem of imbalance of cooperation fairness and model performance in a non-independent identically distributed (Non-IID) data scene. Non-IID data often causes model performance reduction and client contribution evaluation deviation, and an existing method causes inaccurate contribution quantification and unbalanced reward distribution due to single-dimensional evaluation. Therefore, the framework realizes breakthrough through innovative design: firstly, a multi-factor evaluation mechanism (MES) is constructed, client contribution is accurately quantified from data scale, distribution diversity and historical contribution record three dimensions, and single-dimension limitation is avoided; then, through an adaptive dynamic reputation management module, the reputation is updated by adopting a time-varying weight coefficient according to the accuracy of a local model, the data diversity is guaranteed at the initial stage, and the model quality is improved by focusing high-value data at the later stage; and finally, a gradient reward distribution module is constructed based on Jensen-Shannon divergence, the loss distribution difference between the client and the global model is quantified, differential parameter aggregation is realized, linear matching of contribution and rewards is ensured, and fairness is improved. And all the modules cooperate, so that the overall model performance and fairness of the framework are remarkably improved in the environment of high data isomerism and non-uniform category distribution. Experiments show that the method is superior to the prior art on multiple reference data sets, and an efficient and fair solution is provided for federal learning.

Description

Technical Field

[0001] The present invention belongs to the field of federated learning, and particularly relates to a collaborative fair federated learning framework based on multi-dimensional aggregation and adaptive client selection. Background Art

[0002] Federated learning, as an innovative distributed machine learning paradigm, allows multiple data holders to jointly participate in the model training process without directly sharing their local data, thereby achieving the dual goals of data privacy protection and model performance optimization. It is applicable to scenarios where data is scattered, strict privacy protection is required, and fair distribution of model training results is needed. In the framework of federated learning, collaborative fairness is regarded as a key factor for maintaining the healthy operation of the system. By designing a reasonable incentive mechanism, it ensures that each participant can obtain corresponding rewards based on their contributions, thus promoting active participation and long-term cooperation. In practical applications, the datasets of participants often exhibit the characteristics of non-independent and identically distributed (Non-IID), that is, the data distributions are heterogeneous, and there are significant differences in the data scale and quality of each party. Achieving collaborative fairness can improve the performance, generalization ability, and data diversity of the global model, encourage more clients to participate, and enhance the robustness of the model. At the same time, on the premise of protecting data privacy, it ensures that the contributions of all parties are reasonably evaluated and rewarded, realizing the efficient utilization of data value and the balanced distribution of interests.

[0003] The application scenarios of collaborative fair federated learning are extensive, especially suitable for fields with data isolation, strong privacy sensitivity, and the need for multi-party collaboration. For example, in the field of healthcare, multiple hospitals can collaborate to train a disease diagnosis model without sharing patient data, and at the same time, through a fairness mechanism, ensure that hospitals with less data can also obtain reasonable model benefits and avoid being marginalized. In the financial field, multiple banks can jointly build an anti-fraud model, each using its local customer data for training, and at the same time, through a fairness mechanism, balance the contributions of each bank to prevent the data-dominant party from monopolizing the model benefits and promoting the enthusiasm for cooperation. In the field of intelligent Internet of Things, multiple device manufacturers can collaborate to optimize the device performance prediction model, protect the privacy of their respective user data, and at the same time, through a fairness mechanism, ensure that manufacturers with less technical contributions can also share the commercial value brought by the model and promote ecological win-win.

[0004] Currently, federated learning mainly relies on methods based on similarity, data volume, and label diversity to ensure collaborative fairness. However, these methods have significant limitations in practical applications. Firstly, similarity-based methods assign rewards by comparing the similarity between client gradients and the global gradient. But they may overly rely on gradient similarity, resulting in excessive rewards for clients with large amounts of data and ignoring the differences in data quality. Secondly, data volume-based methods allocate rewards based on the client's data volume and model quality. However, it is difficult to comprehensively measure the multi-dimensional contributions of clients, and may underestimate the value of clients with high data quality but small data volume. Finally, although label diversity-based methods consider the uniformity of label distribution, they still cannot accurately evaluate the comprehensive contributions of clients in the case of significant data heterogeneity. These limitations lead to insufficiently fine-grained contribution evaluation, which may cause unfair reward distribution, weaken the enthusiasm of clients to participate, and thus affect the collaborative fairness and long-term stability of the federated learning system. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a federated fairness learning framework based on multi-dimensional aggregation and adaptive client selection, which can effectively balance collaborative fairness and global model performance, ensure collaborative fairness, and obtain competitive prediction accuracy.

[0006] The technical solution of the present invention is as follows:

[0007] (1) The server initializes the global model and the initial reputation scores of all clients, and distributes the global model to all participating clients to start the federated learning process.

[0008] (2) The client trains the global model on its local dataset and uploads the local model update to the server.

[0009] (3) The server calculates a multi-factor evaluation score based on the performance of the local model uploaded by the client and its data characteristics (such as data scale and data diversity), and optimizes the model parameter aggregation strategy based on this.

[0010] (4) The adaptive client selection module selects the clients suitable for participating in the next round of training according to the multi-factor evaluation scores of the clients, ensuring that more contributing clients are selected first. At the same time, a random selection mechanism is introduced to enhance the robustness of the model.

[0011] (5) The dynamic reputation management module dynamically adjusts the reputation scores of all clients according to their training performance (such as the accuracy of the model and data contribution). Among them, the selected clients have a greater chance of increasing their reputation scores, while the unselected clients are more likely to have their reputation scores decreased.

[0012] (6) The dynamic gradient reward allocation module adjusts the gradient contribution ratio according to the multi-factor evaluation scores of each client. The server aggregates the gradient updates of the selected clients through weighted aggregation to ensure that clients with greater contributions have a greater impact on the global model.

[0013] Furthermore, calculating the multi-factor evaluation scores of each client in step (3) includes the following steps:

[0014] 3-1) Based on steps (1) and (2), the server performs weighted synthesis of the multi-dimensional characteristics of the clients through a multi-factor evaluation mechanism to calculate the score of each client. The formula is as follows:

[0015]

[0016] where, w rep , w data , w div are the weight coefficients of reputation, data volume, and data diversity, used to balance the influence of different factors on the final score. Their value range is [0,1], and w rep + w data + w div = 1. represents the reputation score of the client in the previous round of training, n i is the data volume of the client, and d i is a measure of the data diversity of the client.

[0017] The data diversity score d i is calculated through the entropy value of the client data category distribution. The specific formula is:

[0018]

[0019] where, Entropy(N) = -∑ j p j log(p j ), N is the count of each category in the client i's data, and p j is the proportion of category j in the client's dataset.

[0020] Furthermore, the specific process of dynamic client selection in step (4) includes the following steps:

[0021] 4-1) Sort the clients in descending order of MES scores and select the top M high-score clients (M < K, where K is the total number of clients).

[0022] 4-2) Randomly select N clients from the remaining clients (N = Ceil((K - M)*P), where P is the selection ratio, usually 10%) to maintain data diversity.

[0023] 4 - 3) Combine the selected client set C = {Top M} ∪ {Random N s}.

[0024] Furthermore, in step (5), the dynamic reputation management module dynamically adjusts its reputation score according to the training performance of all clients (such as the accuracy of the model and data contribution), including the following steps:

[0025] 5 - 1) For the selected client i, its reputation value will be positively adjusted according to the normalization of the local model accuracy in this round: For the unselected clients, by gradually decaying the reputation, inactive or poorly performing clients are effectively suppressed, enabling the model to gradually focus on the data contributions of high - quality clients:

[0026]

[0027] Among them, α

[0028]

[0029] is the reputation adjustment coefficient, and its update formula is: r

[0030]

[0031] α max is the preset maximum reputation adjustment coefficient, t is the current training round, and T is the total number of training rounds.

[0032]

[0032] Furthermore, in step (6), it includes the following steps:

[0033] 6 - 1) The reward weight calculation formula of the dynamic gradient reward distribution module is:

[0034]

[0035] Among them, β is a hyperparameter used to control the influence degree of reputation on reward distribution, and JS(Loss i , Loss server ) represents the Jensen - Shannon divergence of the loss distribution between client i and the server.

[0036] 6 - 2) The server distributes the aggregated gradient to each client according to the reward weight q i of the client through a sparsification operation. The specific formula is as follows:

[0037]

[0038] 6-3) The server weights and aggregates the client gradients based on the obtained multi-factor evaluation scores. The aggregation update formula for the server to the client gradients is as follows:

[0039]

[0040] where, is the gradient update sent by client i to the server after normalization and scaling:

[0041]

[0042] represents the model gradient update of client i on the local data. γ is a normalization constant used to prevent gradient explosion.

[0043] Compared with the prior art, the advantages of the present invention are as follows:

[0044] (1) By the organic combination of the adaptive client selection module and the dynamic reputation management module, the present invention takes into account fairness while ensuring the model performance, and compensates the unselected clients by introducing a reasonable random strategy, thus improving the generalization performance of the model.

[0045] (2) The dynamic gradient reward allocation module of the present invention realizes a reward allocation mechanism based on Jensen-Shannon divergence, ensuring the consistency of contribution and return. At the same time, the gradient sparsification technology is used to avoid the noise propagation of low-contribution clients, further enhancing the stability and robustness of model updates.

[0046] (3) The multi-dimensional and dynamic evaluation framework of the present invention not only effectively solves the fairness problem in the data heterogeneous environment, but also realizes the synchronous improvement of model performance. In a highly heterogeneous data environment, it can effectively encourage high-quality (i.e., high multi-factor evaluation score) clients to participate, while avoiding the negative impact of low-contribution clients on the model stability.

[0047] (4) The present invention is superior to the existing collaborative fairness methods in terms of improving collaborative fairness and global model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is the overall flowchart of the present invention.

[0049] Figure 2 is the system framework diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0050] The present invention will be further described below in conjunction with embodiments and the accompanying drawings of the specification. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent forms of modification made by those skilled in the art fall within the scope defined by the appended claims of this application.

[0051] See Figure 1 、 Figure 2 , this embodiment discloses a federated fairness learning framework based on multi-dimensional aggregation and adaptive client selection. The method includes the following steps:

[0052] Step 1: The server initializes the global model and the initial reputation scores of all clients, and distributes the global model to all participating clients to start the federated learning process.

[0053] Step 2: The client trains the global model on its local dataset and uploads the updated local model after training to the server.

[0054] Step 3: The server calculates a multi-factor evaluation score based on the performance of the local model uploaded by the client and its data characteristics (such as data scale and data diversity), and optimizes the model parameter aggregation strategy based on this.

[0055] (1) The server performs a weighted synthesis of the multi-dimensional characteristics of the clients participating in the training through a multi-factor evaluation mechanism, calculates the score of each client, and the formula is as follows:

[0056]

[0057] where, w rep , w data , w div are the weight coefficients of reputation, data volume, and data diversity, used to balance the influence of different factors on the final score represents the reputation score of the client in the previous round of training; n i is the data volume of the client, and d i measures the data diversity of the client.

[0058] The data diversity score d i is calculated through the entropy value of the client data category distribution, and the specific formula is:

[0059]

[0060] where, Entropy(N) = -∑ j p j log(p j ), N is the count of each category in the client i data, p jis the proportion of category j in the client dataset.

[0061] Step 4: The adaptive client selection module selects the clients suitable for participating in the next round of training based on the multi-factor evaluation scores of the clients, ensuring that more contributing clients are selected first. At the same time, a random selection mechanism is introduced to enhance the robustness of the model.

[0062] (1) Sort the clients participating in the training in descending order of MES scores and select the top M high-scoring clients (M < K, where K is the total number of clients).

[0063] (2) Randomly select N clients (N = Ceil((K - M) * P), where P is the selection ratio, usually 10%) from the remaining clients to maintain data diversity.

[0064] (3) Combine the selected client set C = {Top M} ∪ {Random N s}.

[0065] Step 5: The dynamic reputation management module dynamically adjusts the reputation scores of all clients according to their training performance (such as the accuracy of the model and data contribution). Among them, the selected clients have a greater chance of increasing their reputation scores, while the unselected clients are more likely to have their reputation scores decreased.

[0066] (1) For the selected client i, its reputation value will be positively adjusted according to the normalization of the local model accuracy in this round:

[0067]

[0068] (2) For the unselected clients, by gradually decaying the reputation, inactive or poorly performing clients are effectively suppressed, enabling the model to gradually focus on the data contributions of high-quality clients:

[0069]

[0070] (3) Among them, α r is the reputation adjustment coefficient, and its update formula is:

[0071]

[0072] α max is the preset maximum reputation adjustment coefficient, t is the current training round, and T is the total number of training rounds. This formula gradually increases α through linear interpolation r to ensure that as the training progresses, the adjustment strength of the reputation gradually increases, so that in the later stage of training, it can focus more on the long-term performance of high-quality clients.

[0073] Step 6: The dynamic gradient reward allocation module adjusts the gradient contribution ratio according to the multi-factor evaluation scores of each client. The server ensures that clients with greater contributions have a greater impact on the global model by weighted aggregation of the selected clients' gradient updates.

[0074] (1) The server performs weighted aggregation on the client gradients based on the obtained multi-factor evaluation scores. The aggregation update of the client gradients by the server

[0075] The formula is:

[0076]

[0077] where is the gradient update sent by client i to the server after normalization and scaling:

[0078]

[0079] represents the model gradient update of client i on local data, and γ is a normalization constant used to prevent gradient explosion.

[0080] (2) The formula for calculating the reward weight of the dynamic gradient reward allocation module is:

[0081]

[0082] where β is a hyperparameter used to control the influence degree of reputation on reward allocation, and JS(Loss i , Loss server ) represents the Jensen-Shannon divergence of the loss distribution between client i and the server.

[0083] In this way, the dynamic gradient reward allocation module ensures that clients with higher reputation and more similar data distribution to the server will obtain a greater reward weight q i .

[0084] (3) Next, the server allocates the aggregated gradient to each client according to the reward weight q i of the client through a sparsification operation. The specific formula is as follows:

[0085]

[0086] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A federated fairness learning framework based on multi-dimensional aggregation and adaptive client selection, characterized in that It includes the following steps: Step 1: The server initializes the global model and the initial reputation scores of all clients, and distributes the global model to all participating clients; Step 2: The clients use local data to train the global model and upload the updated local models to the server; Step 3: The server calculates the multi-factor evaluation score (MES) of each client based on the performance of the client's local model, the data volume, and the data diversity; Step 4: The adaptive client selection module selects the clients suitable for participating in the next round according to the MES scores of the clients; Step 5: The dynamic reputation management module dynamically adjusts the reputation scores of the clients according to their training performance; Step 6: The dynamic gradient reward allocation module adjusts the gradient contribution ratio of each client according to its MES, and the server updates the global model through weighted aggregation.

2. The federated fairness learning framework based on multi-dimensional aggregation and adaptive client selection according to claim 1, characterized in that, The calculation formula for the multi-factor evaluation score (MES) in Step 3 is: Among them, w rep , w data , w div are the weight coefficients of reputation, data volume, and data diversity, used to balance the impact of different factors on the final score. Their value range is [0, 1], and w rep + w data + w div = 1; represents the reputation score of the client in the previous round of training, n i is the data volume of the client, d i then measures the data diversity of the client; the said d i is the data diversity score, calculated through the entropy value of the client data category distribution. The specific formula is: Among them, Entropy(N) = -∑ j p j log(p j ), N is the count of each category in the data of client i, and p j is the proportion of category j in the client dataset.

3. A federated fairness learning framework based on multi-dimensional aggregation and adaptive client selection according to claim 1, characterized in that In Step 4, the specific process of dynamic client selection is as follows: 3-1) Sort the clients in descending order of MES scores, and select the top M high-score clients (M < K, where K is the total number of clients); 3-2) Randomly select N clients (N = Ceil((K - M) * P), where P is the selection ratio, usually 10%) from the remaining clients to maintain data diversity; 3-3) Merge the selected client set C = {Top M} ∪ {Random N s}.

4. A federated fairness learning framework (FedMDA) based on multi-dimensional aggregation and adaptive client selection according to claim 1, characterized in that, In Step 5, the reputation update formula of the dynamic reputation management (DRM) module is: For the selected client i, its reputation value will be adjusted positively according to the normalization of the local model accuracy in this round as follows: For the clients not selected, by gradually decaying the reputation, inactive or poorly performing clients are effectively inhibited, enabling the model to gradually focus on the data contributions of high-quality clients: where α r is the reputation adjustment coefficient, and its update formula is: α max is the preset maximum reputation adjustment coefficient, t is the current training round, and T is the total number of training rounds. This formula gradually increases α through linear interpolation r to ensure that as the training progresses, the adjustment of reputation gradually increases, so that in the later stage of training, more focus can be placed on the long-term performance of high-quality clients.

5. A federated fairness learning framework based on multi-dimensional aggregation and adaptive client selection according to claim 1, characterized in that, In Step 5, the server performs weighted aggregation on the client gradients according to the obtained multi-factor evaluation scores. The aggregation update formula for the server on the client gradients is: wherein, is the gradient update sent by client i to the server after normalization and scaling: It represents the model gradient update of client i on local data, and γ is a normalization constant used to prevent gradient explosion.

6. A federated fairness learning framework based on multi-dimensional aggregation and adaptive client selection according to claim 1, characterized in that, Step 6 includes the following steps: (6-1) The reward weight calculation formula of the dynamic gradient reward allocation (DGR) module is: Among them, β is a hyperparameter used to control the degree of influence of reputation on reward distribution, and JS(Loss i , Loss server ) represents the Jensen-Shannon divergence of the loss distribution between client i and the server. In this way, the dynamic gradient reward allocation module ensures that clients with higher reputation and more similar data distribution to the server will receive greater reward weights (6-2) Next, the server sparsifies the aggregated gradients and assigns them to each client according to the reward weight q of the client i using the following specific formula:

7. A federated learning system, characterized in that, It includes the method according to any one of claims 1-6, which is used to achieve the following goals in a non-independent and identically distributed (Non-IID) data environment: ● Encourage high-quality clients to participate through dynamic reputation management; ● Balance data diversity through multi-dimensional evaluation and random selection; ● Improve model performance and cooperation fairness based on gradient reward allocation.

Images