A Federated Learning Method and Device Based on Social Grouping

Abstract

The present invention discloses a federated learning method and device based on social grouping, which conducts social trust evaluation based on direct social trust and indirect social trust; users participating in the federated learning task form a plurality of social groups that are Nash stable and non-overlapping according to the social trust values; members of the social group use local data to train and obtain local model parameter updates, and determine a Gaussian noise perturbation strategy according to the social trust value with the manager of the social group to obtain the perturbed local model parameter updates; the manager of the social group aggregates the local model parameter updates of all members within the social group to obtain a social layer pre-aggregated model parameter update, and the global aggregator aggregates the social layer pre-aggregated model parameter updates of all social groups to obtain a global aggregated model. The present invention can improve the usability of the federated learning service while satisfying customized privacy protection, resist free-riding attacks, and achieve efficient and privacy-protected distributed Internet of Things data sharing.

Description

Technical Field

[0001] The present invention belongs to the field of information security, and particularly relates to a federated learning method and device based on social grouping. Background Art

[0002] With the explosive growth of intelligent Internet of Things (IoT) devices such as smart phones, wearable devices, and connected vehicles, a vast amount of IoT data will be generated, collected, and processed at distributed terminal devices. Due to issues such as data ownership rights, asset recognition, and user privacy, current IoT big data presents characteristics such as data "island" and knowledge "fragmentation". How to effectively aggregate and share the scattered IoT data, thereby mining the value in the data and providing various personalized and intelligent IoT intelligent applications and services has become an important requirement. However, traditional deep learning needs to gather a vast amount of scattered IoT data to a central cloud computing node for data mining and knowledge extraction, which may lead to serious privacy leakage and data abuse.

[0003] Federated Learning, as a new machine learning paradigm for privacy protection, is a key technology to break through the data island dilemma, achieve secure and efficient sharing of large-scale IoT data, and fully release the value of IoT big data. In federated learning, data owners use local data to periodically train local models and send them to an aggregation server for global aggregation to synthesize a global model, which is then sent back to the data owners for the next round of training. Since data owners only share the model parameters learned from local data rather than the original data, the original data does not leave the domain and the data is available but invisible. However, a large number of studies and experiments have found that for the local model parameters (such as gradients) shared by users in federated learning, attackers may still restore the users' privacy and sensitive information by launching advanced attacks such as model inversion and membership inference. Due to the advantages of differential privacy mechanism such as strict theoretical guarantee and low computational overhead, existing security countermeasures are mainly implemented based on Differential Privacy technology. However, although a large differential privacy noise can provide strong privacy protection, it will lead to a serious decline in model performance. Therefore, it is urgent to balance privacy and practicality. In addition, IoT users usually have different privacy protection requirements, and there are generally free-riding users in federated learning, resulting in a reduction in the quality of the federated learning model. Current federated learning solutions are difficult to provide high-availability federated learning services that meet customized privacy protection and resist free-riding attacks. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention provides a federated learning method and device based on social grouping, which can improve the availability of federated learning services while meeting customized privacy protection, resist free-riding attacks, promote the fairness of federated learning reward distribution, and achieve efficient and privacy-protected distributed Internet of Things data sharing.

[0005] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0006] A federated learning method based on social grouping, comprising:

[0007] For users jointly participating in a federated learning task, construct a social relationship graph among the users, and each user in the social relationship graph evaluates the social trust value of other users based on direct social trust and indirect social trust;

[0008] The users jointly participating in the federated learning task form multiple social groups that are Nash stable and non-overlapping according to the social trust value;

[0009] All members within each social group use local data for training to obtain local model parameter updates, and determine a Gaussian noise perturbation strategy according to the social trust value with the manager of the social group to obtain the perturbed local model parameter updates. The manager of the social group is the user with the highest centrality within the social group;

[0010] The manager of each social group performs in-group model pre-aggregation on the local model parameter updates of all members within the social group to obtain social layer pre-aggregated model parameter updates;

[0011] Each social group transmits the social layer pre-aggregated model parameter updates to the global aggregator to obtain a global aggregated model, and the global aggregator transmits the global aggregated model to each user participating in the federated learning task.

[0012] Further, the users jointly participating in the federated learning task form multiple social groups that are Nash stable and non-overlapping according to the social trust value, specifically:

[0013] For each social group, define a group utility function according to the social trust value between users;

[0014] For each social group, design a fair distribution rule according to the group utility function to obtain the individual utility functions of all members within the social group;

[0015] For each user participating in the federated learning task, construct a transferable group list according to the individual utility function, and send a join application to the optimal social group in the transferable group list while satisfying the social group departure rule;

[0016] For each social group, when receiving multiple join applications, construct an acceptable user list, and accept the join application of the optimal user in the acceptable user list while satisfying the social group join rule, and reject the join applications of other users in the acceptable user list;

[0017] Perform user separation and merger operations, and update the set of current social groups;

[0018] Repeat the above steps until the state of the social groups no longer changes, forming the multiple social groups that are Nash stable and disjoint.

[0019] Furthermore, the definition of the group utility function according to the social trust value between users is specifically as follows:

[0020] The group utility of the social group Φ j is the difference between the benefit and the cost:

[0021]

[0022] In the formula, λ p is the reward payment for the unit model quality; q n is the local model quality of user n in the social group Φ j , q n is related to the social relationship between user n and the manager of the social group Φ j and the non-i.i.d. degree of user n's local data; λ c is a proportionality coefficient greater than 0; |Φ j | represents the size of the social group Φ j , that is, the number of members in the social group; λ c |Φ j | represents the communication cost within the social group Φ j ; represents the model quality when user n forms a single group;

[0023] The local model quality q j of user n in the social group Φ n is negatively correlated with the model loss value , specifically as follows:

[0024]

[0025] where κ1 and κ2 are both positive curve fitting parameters; κ2 represents the asymptotically maximum local model quality when the model loss approaches 0;

[0026] the social grouping Φ j the local model loss of user n is related to the Gaussian noise size σ n，j and the degree of non-independent and identically distributed data γ, specifically:

[0027]

[0028] where μ1, …, μ5 are all positive curve fitting parameters; the numerator part μ1exp(-μ2·γ) reflects the decrease in the marginal loss value of the local model when the degree of non-independent and identically distributed data γ increases; the denominator part μ s +exp(-μ4·σ n，j ) reflects the decrease in the local model quality when the model noise size σ n，j increases.

[0029] Furthermore, designing a fair distribution rule according to the group utility function to obtain the individual utility functions of all members within the social grouping, specifically:

[0030] the social grouping Φ j the individual utility of each member user n within is the sum of the utility in the non-cooperative case and the additional utility distributed according to the contribution ratio:

[0031]

[0032] where and respectively represent the non-cooperative utilities of users l and n, that is, the group utilities when users l and n form single-item groupings respectively; represents the contribution weight of user n in the social grouping Φ j ; represents the additional utility generated when all members within the social grouping Φ j cooperate compared to the non-cooperative case.

[0033] Furthermore, constructing a transferable grouping list according to the individual utility function, specifically:

[0034] Constructing a preference order according to the individual utility function;

[0035] Constructing a transferable grouping list according to the preference order;

[0036] Constructing a preference order according to the individual utility function, specifically:

[0037] user n, n ∈ Φj For the social grouping Φ′ j , Φ′ j ≠Φ j The preference order is as follows:

[0038]

[0039] Where ψ n (Φ′ j ∪{n}) represents the individual utility of user n after joining the social grouping Φ′ j ; ψ l (Φ′ j ) and ψ l (Φ′ j ∪{n}) respectively represent the individual utilities of user l∈Φ′ j in the original social grouping Φ′ j and the new social grouping Φ′ j ∪{n}; represents the set of social groupings that have historically rejected the join application of user n;

[0040] According to the above preference order, constructing a transferable grouping list is specifically as follows:

[0041] At the t-th iteration, the transferable grouping list of user n in the social grouping Φ j is:

[0042]

[0043] Where Φ (t) represents the set of all social groupings at the t-th iteration.

[0044] Furthermore, the social grouping departure rule is specifically as follows:

[0045] If the social grouping Φ j accepts the join application of a new user at the t-th iteration, then none of the original internal members can leave the social grouping Φ j at this t-th iteration;

[0046] The optimal social grouping in the transferable grouping list is specifically as follows:

[0047] For user n, n∈Φ j , the optimal social grouping in the transferable grouping list is If then user n prefers to leave the original social grouping and form a single-item grouping; otherwise, user n prefers to leave the original social grouping and join the social grouping

[0048] Furthermore, the social grouping joining rule is specifically as follows:

[0049] If there are internal members leaving the social grouping in the t-th iteration, then the social grouping cannot accept the joining applications of all new users in the t-th iteration;

[0050] The optimal user in the acceptable user list is specifically as follows:

[0051] For social grouping the optimal user in the acceptable user list is n * = arg max ρ n (Φ j ), where is the set of acceptable users of social grouping Φ j at the t-th iteration, that is, the set of users who send joining applications to social grouping Φ j at the t-th iteration.

[0052] Furthermore, the method for determining the Gaussian noise perturbation strategy according to the social trust value with the manager of the social grouping is specifically as follows:

[0053] If the social trust value between user n and the manager of social grouping Φ j is greater than the system preset threshold, then directly transmit the updated original local model parameters to the manager of social grouping Φ j ;

[0054] If the social trust value between user n and the manager of social grouping Φ j is lower than the system preset threshold, then transmit the updated local model parameters with Gaussian noise perturbation to the manager of social grouping Φ j ;

[0055] If user n forms a single-item grouping, then add the maximum Gaussian noise perturbation with a size of σ max to the updated local model parameters, and user n is the manager of the single-item grouping.

[0056] Furthermore, the updated local model parameters with Gaussian noise perturbation are specifically as follows:

[0057] Each member within each social grouping maps the social trust value with the manager of the social grouping to a privacy protection level;

[0058] Each member within each social grouping calculates the size parameter of the Gaussian noise according to the privacy protection level;

[0059] Each member within each social group adds corresponding Gaussian noise to the local model parameter update according to the magnitude parameter of the Gaussian noise, obtaining the perturbed local model parameter update;

[0060] Mapping the social trust value with the manager of the social group to a privacy protection level specifically as follows:

[0061]

[0062] In the formula, θ1 and θ2 are positive adjustment parameters; ∈ n，j Denotes the privacy budget, the smaller its value, the greater the degree of privacy protection; α n，j Denotes the social trust value between user n and the manager of social group Φ j ;

[0063] Each member within each social group calculates the magnitude parameter of the Gaussian noise according to the privacy protection level, specifically as follows:

[0064]

[0065] In the formula, δ is a small probability value; σ n，j Is the variance value of the Gaussian noise, the larger its value, the greater the magnitude of the added noise;

[0066] σ n，j ∈[0, σ max , σ max Denotes the maximum value of the Gaussian noise variance.

[0067] A federated learning device based on social groups, comprising:

[0068] A construction module, configured to construct a social relationship graph among users for users jointly participating in a federated learning task, where each user in the social relationship graph evaluates the social trust value of other users based on direct social trust and indirect social trust;

[0069] A social group formation module, configured to enable the users jointly participating in the federated learning task to form a plurality of Nash-stable and non-overlapping social groups according to the social trust value;

[0070] A local model parameter update module, configured to enable all members within each social group to use local data for training to obtain local model parameter updates, and determine a Gaussian noise perturbation strategy according to the social trust value with the manager of the social group, obtaining the perturbed local model parameter updates, where the manager of the social group is the user with the largest centrality within the social group;

[0071] A pre-aggregation module, which is used for the manager of each social group to perform pre-aggregation of models within the social group on the updates of the local model parameters of all members within the social group, so as to obtain the updates of the pre-aggregated model parameters at the social layer;

[0072] A transmission module, which is used for each social group to transmit the updates of the pre-aggregated model parameters at the social layer to a global aggregator to obtain a global aggregated model, and the global aggregator transmits the global aggregated model to each user participating in the federated learning task.

[0073] Compared with the prior art, the present invention has at least the following beneficial effects:

[0074] (1) Compared with the existing mainstream centralized machine learning mode, the present invention is based on the federated learning mode. The local data of data owners is retained in the terminal device, and only the updates of the local model parameters obtained by training from the local data need to be periodically sent to the global aggregator to perform the global model aggregation operation, realizing the distributed data sharing and collaborative model training of Internet of Things terminal devices, and solving the technical problem of explicit privacy data leakage existing in the large-scale data sharing of the existing Internet of Things.

[0075] (2) The present invention proposes a customized privacy protection method based on social trust. Aiming at the diverse privacy protection needs of different Internet of Things users under different federated learning services, the social trust between users is comprehensively evaluated through direct social trust evaluation and indirect social trust evaluation, and the mapping relationship between social trust and privacy protection degree is proposed to realize personalized privacy protection in federated learning.

[0076] (3) The present invention proposes a highly available federated learning method with social awareness. Combining the social attributes, competition and cooperation characteristics between users, a social grouping layer is introduced into the traditional two-layer federated learning architecture including a centralized aggregator and distributed terminal devices, a new security hypothesis is defined by using social trust, and appropriate perturbation addition and model pre-aggregation operations are performed within the social group according to social trust, improving the quality and performance of the federated learning model.

[0077] (4) The present invention proposes a method for fair distribution of user benefits against free-riding attacks. In the case of free-riding attacks, by evaluating the size of the user's local data, the degree of non-independent and identically distributed, and the size of the noise, the quality of the user's local model is evaluated and the user's contribution is calculated, and the learning rewards are fairly distributed according to the user's contribution value, so as to provide accurate user participation incentives to suppress the user's free-riding behavior while motivating the user's high-quality model training.

[0078] (5) The present invention proposes a method for forming dynamic social groups based on Nash stability in game cooperation. By designing a group utility function, an individual utility function, and a bilateral dynamic matching process between users and social groups, the optimal strategies of users and social groups in a highly dynamic environment are obtained, enhancing the efficiency of social grouping in federated learning and adapting to the dynamic withdrawal and joining of users in federated learning.

[0079] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following specifically gives preferred embodiments and, in conjunction with the accompanying drawings, makes a detailed description as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the following will briefly introduce the drawings required for use in the description of the specific embodiments. Obviously, the following-described drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0081] Figure 1 It is a flowchart of the federated learning method based on social grouping of the present invention;

[0082] Figure 2 It is an architecture diagram of social perception-based grouped federated learning;

[0083] Figure 3 It is a flowchart of forming a dynamic optimal social group structure with Nash stability;

[0084] Figure 4 It is a flowchart of the noise perturbation strategy decision of users in the optimal group;

[0085] Figure 5 It is a flowchart of distributed model training and hierarchical model aggregation in federated learning. SPECIFIC EMBODIMENTS

[0086] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0087] As a specific embodiment of the present invention, as Figure 1 shown, a federated learning method based on social grouping of the present invention specifically includes the following steps:

[0088] Step 1. For users jointly participating in the federated learning task, construct a social relationship graph among users. In the social relationship graph, each user evaluates the social trust value of other users based on direct social trust and indirect social trust.

[0089] That is, for a group of users jointly participating in the federated learning task, according to the social relationship among users, each user evaluates the social trust value of existing and potential social friends.

[0090] Specifically, the social perception-based grouped federated learning architecture is as Figure 2 shown, including the terminal layer, the social network, and the cloud layer. The cloud server, as the global aggregation server in federated learning, is assumed to be an honest but curious entity. That is, the cloud server will honestly perform the global model aggregation operation in each global communication round k, but will infer the privacy information in the local model updates of users through data analysis and other means. The terminal layer consists of N Internet of Things (IoT) users jointly participating in the federated learning task, and its set is IoT users with different social attributes are interconnected through the social network. The N users jointly participating in the federated learning task can dynamically form multiple disjoint social groups, and its set is represented as Φ = {Φ1,..., Φ j ,..., Φ J}.

[0091] In a global communication round k during the federated learning process, first, each IoT user uses its intelligent terminal device (such as a smartphone and a connected vehicle) to jointly train a globally shared artificial intelligence model on local private data. Among them, the IoT user only needs to send the local model parameter updates obtained from local data training according to the downloaded global model. Then, for each social group Φ j ∈Φ, the manager φ j of this group performs a social layer pre-aggregation operation on the local model parameter updates of all users in the group and transmits the pre-aggregation result to the global aggregator. Then, the global aggregator aggregates the social layer pre-aggregation model parameters of all social groups to obtain a global model and sends it back to all IoT users for the next round of training. Iterate the above process until the model reaches the ideal performance index or the maximum number of global communication rounds is reached.

[0092] As a preferred implementation manner, each user in the social relationship graph evaluates the social trust value of other users based on direct social trust and indirect social trust, that is, evaluates the social trust value according to the social relationship among users, specifically including:

[0093] Step 1.1. Evaluate the direct trust value among users according to the social relationship among users;

[0094] Specifically, let represent the social graph among the IoT users in the set . Among them, is the edge set among the IoT users. For en,m ∈ [0, 1], it represents the social relationship, that is, the social intimacy, between user n and user m (n ≠ m). Among them, e n，m = e m，n . e n，m = 1 indicates that the social relationship between the two users is the strongest, and e n，m = 0 indicates that the two users are strangers. Let α n，m ∈ [0, 1] represent the social trust value between two users n and m, and its set is denoted as The social trust value α n,m is related to the direct social interaction, the topological relationship of the social graph, and the time decay effect between user n and user m. Among them, the direct social trust e n,m comes from the historical interaction (such as sharing Weibo, photos, and videos) experience between user n and user m, and is affected by the duration of each interaction and the time of interaction:

[0095]

[0096] In the formula, and are the total numbers of positive and negative interactions between user n and user m respectively; ν > 0 is the penalty factor; is the duration of the b-th interaction; Dth is the threshold of the duration. If then Otherwise, Keep the original value; describes the exponential time decay effect; t b is the occurrence time of the b-th interaction; is the decay rate.

[0097] Step 1.2: Evaluate the indirect trust value between users according to the direct trust value;

[0098] Specifically, since the number of direct interactions between IoT users is small, it is necessary to comprehensively evaluate the social trust between users by combining the topological relationship in the social graph. Let T path represent the shortest path connecting user n and user m in the social graph . Among them, |T path | is called the social distance. The indirect social trust value τ n，m between user n and user m can be calculated as the aggregated recommended trust of the social friends of user n on user m on the path T path , that is,

[0099]

[0100] In the formula, represents the path T of users l and k in the social graph and are adjacent on it. path

[0101] Step 1.3: Obtain the social trust value according to the direct trust value and the indirect trust value;

[0102] Specifically, the social trust value α between user n and user m n,m is the weighted average of the direct social trust value and the indirect social trust value:

[0103] α n,m = ωe n,m +(1 - ω)τ n，m

[0104] In the formula, w ∈ [0, 1] is the weight factor. Among them, α n,m ∈ [0, 1]. Denote γ n，j as the centrality of user n in the social cluster Φ j , that is, the number of neighbors of user n in the cluster Φ j , and its value is where f n，l = {0, 1}, if e n，l > 0, then f n，l = 1; otherwise, f n，l = 0.

[0105] Step 2: The users jointly participating in the federated learning task form multiple social groups that are Nash stable and non - overlapping according to the social trust value, as Figure 3 described, specifically including:

[0106] Step 2.1: Initialize the social groups;

[0107] Specifically, at the initial moment (i.e., t = 0), set the social group state to Φ (0) , and this initial state depends on the specific application (such as the stable partition result of the previous federated learning task).

[0108] Step 2.2: When t ≥ 1, for each social group, define the group utility function according to the social trust value between users;

[0109] Specifically, the group utility of the social group Φ j is the difference between the benefit and the cost:

[0110]

[0111] In the formula, λ p ​is the reward payment for the quality of the unit model; q n is the social group Φ j the local model quality of user n in n is related to the social relationship between user n and the social group Φ j and the manager, as well as the degree of non - independent and identically distributed of user n's local data; λ c is a proportionality coefficient greater than 0; |Φ j | represents the size of the social group Φ j that is, the number of members within the social group; λ c |Φ j | represents the communication cost within the social group Φ j ; represents the model quality when user n forms a single - item group (i.e., Φ j ={n});

[0112] The local model quality q of user n in the social group Φ j is negatively correlated with the model loss value n Specifically:

[0113]

[0114] In the formula, κ1 and κ2 are both positive curve - fitting parameters; κ2 represents the asymptotic maximum local model quality when the model loss approaches 0;

[0115] The Dirichlet distribution is used to characterize the heterogeneity of data distribution among IoT users. For a classification task of Y classes, the training samples of each IoT user follow a Dirichlet distribution parameterized by the vector a~Dir(γ), and the probability density function of this distribution is:

[0116]

[0117] In the formula, the denominator is the normalization constant; γ y >0, is the degree of non - independent and identically distributed of the data. If γ y →∞, then the data distribution of all users is independent and identically distributed; if γ y →0, then each user only randomly holds one type of data sample. In the present invention, γ y is set to γ,

[0118] The local model loss of user n in the social group Φ j is related to the size of the Gaussian noise σ n，j ​​It is related to the degree of non - independent and identically distributed of the data γ, specifically as follows:

[0119]

[0120] In the formula, μ1, …, μ5 are all positive curve - fitting parameters; the numerator part μ1exp(-μ2·γ) reflects the decrease in the marginal loss value of the local model when the degree of non - independent and identically distributed of the data γ increases; the denominator part μ3 + exp(-μ4·σ n，j ) reflects the decrease in the quality of the local model when the model noise size σ n，j increases.

[0121] Step 2.3: For each social group, according to the group utility function, design a fair distribution rule to obtain the individual utility functions of all members within the social group;

[0122] Specifically, for each member user n within the social group Φ j the individual utility is the sum of the proportional - allocated additional utility and the non - cooperative utility:

[0123]

[0124] In the formula, and respectively represent the non - cooperative utilities of users l and n, that is, the group utilities when users l and n form single - item groups respectively; represents the contribution weight of user n in the social group Φ j ; represents the additional utility generated when all members within the social group Φ j cooperate compared to the non - cooperative situation.

[0125] Step 2.4: For each user participating in the federated learning task, construct a transferable group list according to the individual utility function, and send a join application to the optimal social group in the transferable group list while satisfying the social group leaving rule;

[0126] Specifically, in the present invention, the individual utility function constructs a transferable group list and sends a join application to the optimal social group in the transferable group list while satisfying the social group leaving rule, specifically including:

[0127] Step 2.4.1: Construct a preference order according to the individual utility function;

[0128] User n, n ∈ Φ j has a preference order for the social group Φ′ j , Φ′ j ≠Φ j as follows:

[0129]

[0130] where ψ n (Φ′ j ∪{n}) represents the individual utility of user n after joining the social group Φ′ j ; ψ l (Φ′ j ) and ψ l (Φ′ j ∪{n}) respectively represent the individual utilities of user l∈Φ′ j in the original social group Φ′ j and the new social group Φ′ j ∪{n}; represents the set of social groups that have historically rejected the joining application of user n;

[0131] Step 2.4.2, construct a transferable group list according to the preference order;

[0132] At the t-th iteration, the transferable group list of user n in the social group Φ j is:

[0133]

[0134] where Φ (t) represents the set of all social groups at the t-th iteration;

[0135] Step 2.4.3, design a social group leaving rule;

[0136] The social group leaving rule in the present invention is specifically:

[0137] If the social group accepts the joining application of a new user at the t-th iteration, none of the original internal members can leave the social group at the t-th iteration;

[0138] Step 2.4.4, determine the optimal transfer strategy of the user by calculating the optimal social group;

[0139] For user n, n∈Φ j , the optimal social group in the transferable group list is If , then user n tends to leave the original social group and form a single-group; otherwise, user n tends to leave the original social group and join the social group

[0140] Step 2.5. For each social group, when multiple said joining applications are received, construct an acceptable user list, and while meeting the social group joining rules, accept the joining application of the optimal user in the acceptable user list, and at the same time reject the joining applications of other users in the acceptable user list;

[0141] Step 2.5.1. When multiple said joining applications are received, construct an acceptable user list;

[0142] Let be the set of acceptable users of the social group Φ j at the t-th iteration, that is, the set of users who send joining applications to the social group Φ j at the t-th iteration;

[0143] Step 2.5.2. Design social group joining rules;

[0144] The social group joining rules in the present invention are specifically as follows:

[0145] If there are internal members leaving the social group at the t-th iteration, then the social group cannot accept the joining applications of all new users at the t-th iteration;

[0146] Step 2.5.3. Determine the optimal permission strategy of the social group by calculating the optimal user;

[0147] The optimal user in the acceptable user list of the social group Φ j is n * = arg maxρ n (Φ j ),

[0148] Step 2.6. Execute user separation and merging operations, and update the set of current social groups;

[0149] First, the separation and merging operations of transferring the user n * ∈Φ j to another social group Φ′ j include: separation operation (i.e., ) and merging operation (i.e., ). Among them, and Φ′ j + = Φ′ j ∪{n *};

[0150] Second, update the set of current social groups Φ (t) →Φ (t+1) .

[0151] Step 2.7: Repeat the above steps until the state of the social grouping no longer changes, i.e., a Nash-stable and non-overlapping social grouping Φ is achieved. * 。

[0152] Step 3: All members within each social grouping use local data to train and obtain updated local model parameters, and determine a Gaussian noise perturbation strategy based on the social trust value with the manager of the social grouping. The manager of the social grouping is the user with the highest centrality within the social grouping. Specifically, it includes: Figure 4 As described above.

[0153] Step 3.1: Use local data to train and obtain updated local model parameters.

[0154] Each IoT user n trains the global model Θ received in the previous communication round k - 1 on the local private dataset using the stochastic gradient descent algorithm k-1 and generates the updated local model parameters for the current communication round

[0155]

[0156] where η is the learning rate, is the loss function on the local data samples of user n.

[0157] Step 3.2: Determine the Gaussian noise perturbation strategy.

[0158] Case 1: If the social trust value between user n within the social grouping Φ j and the manager φ j of the social grouping Φ j is greater than the system preset threshold (i.e., α n，j ≥ α th ), then directly transmit the original updated local model parameters (i.e., set the noise magnitude to σ n，j = 0) to the manager of the social grouping Φ j .

[0159] Case 2: If the social trust value between user n within the social grouping Φ j and the manager φ j of the social grouping Φ j is lower than the system preset threshold (i.e., 0 < α n，j < α th ), then transmit the updated local model parameters with appropriate Gaussian noise perturbation to the manager of the social grouping Φ j , where S represents the sensitivity of the L2 norm. Specifically, it includes:

[0160] 1) Each member within each social group maps the social trust value between themselves and the administrator of the social group to a privacy protection level, specifically as follows:

[0161]

[0162] In the formula, θ1 and θ2 are positive adjustment parameters; ∈n ，j represents the privacy budget, and the smaller its value, the greater the degree of privacy protection; α n，j represents the social trust value between user n and the administrator of social group Φ j ;

[0163] 2) Calculate the magnitude parameter of the Gaussian noise according to the privacy protection level, specifically as follows:

[0164]

[0165] In the formula, δ is a small probability value; σ n，j is the variance value of the Gaussian noise, and the larger its value, the greater the magnitude of the added noise;

[0166] σ n，j ∈[0, σ max , σ max represents the maximum value of the Gaussian noise variance.

[0167] Case 3: If user n forms a single-item group (i.e., Φ j ={n}), then add a Gaussian noise perturbation with a maximum magnitude of σ max to the local model parameter update, and user n is the administrator of the single-item group (i.e., φ j =n);

[0168] Step 3.3: Calculate the perturbed local model parameter update;

[0169] By adding the corresponding Gaussian noise perturbation to the local model parameter update, user n generates the perturbed local model parameter update:

[0170]

[0171] Step 4: After completing the local model parameter update and perturbation on the user side, perform hierarchical aggregation of the federated learning model, as Figure 5 described, specifically including:

[0172] Step 4.1: The administrator within each social group performs pre-aggregation of the model within the social group on the local model parameter updates of all members within the social group to obtain the social layer pre-aggregated model parameter update.

[0173] Specifically, each social group Φ jThe manager φ j Aggregate the local model updates of all members within the social grouping and generate the social layer pre-aggregated model parameter updates:

[0174]

[0175] Wherein, is the set of users who send the original local model parameter updates to φ j ; is the set of users who send the perturbed local model parameter updates to φ j ; q n，j is the quality of the local model parameter update of user n within the social grouping Φ j .

[0176] Step 4.2: Each social grouping transmits the social layer pre-aggregated model parameter updates to the global aggregator to obtain a global aggregated model, and the global aggregator transmits the global aggregated model to each user participating in the federated learning task.

[0177] Specifically, the cloud global aggregator aggregates the social layer pre-aggregated model parameter updates uploaded by all social groupings to generate the current global model Θ k :

[0178]

[0179] Wherein, Φ * is the set of Nash-stable social groupings.

[0180] Step 4.3: Determine whether the trained global model Θ k has converged, or whether the current global communication round k has reached the maximum number of iterations. If one of the conditions is met, the model learning stops.

[0181] The present invention provides a federated learning method based on social groupings. By considering the inherent and persistent social associations between users in federated learning, users who trust each other can form a stable social grouping, and appropriate differential privacy noise is added to the original local model parameters according to the strength of the social relationship. Then, the local model parameters of all users within the grouping perform a grouped pre-aggregation operation, thereby preventing the sharp decline in the performance of the federated learning model caused by adding too much differential privacy noise (i.e., improving usability). At the same time, it is difficult for other groups and attackers to reverse-infer the local model parameter information of each user from the pre-aggregated model parameter information, thereby realizing the privacy protection of users within the grouping. The present invention designs a new type of highly available grouped federated learning paradigm that supports customized privacy protection by utilizing the inherent and persistent social attributes between users, thereby improving the usability of the federated learning model while meeting the customized user privacy protection.

[0182] The present invention also provides a federated learning device based on social grouping, including:

[0183] A construction module, configured to construct a social relationship graph among users who jointly participate in a federated learning task, where each user in the social relationship graph evaluates a social trust value for other users based on direct social trust and indirect social trust;

[0184] A social grouping formation module, configured to enable the users who jointly participate in the federated learning task to form a plurality of Nash-stable and non-overlapping social groupings according to the social trust value;

[0185] A local model parameter update module, configured to enable all members within each social grouping to use local data for training to obtain local model parameter updates, and determine a Gaussian noise perturbation strategy according to the social trust value with the administrator of the social grouping, so as to obtain perturbed local model parameter updates, where the administrator of the social grouping is the user with the largest centrality within the social grouping;

[0186] A pre-aggregation module, configured to enable the administrator of each social grouping to perform in-group model pre-aggregation on the local model parameter updates of all members within the social grouping to obtain in-group social layer pre-aggregated model parameter updates;

[0187] A transmission module, configured to enable each social grouping to transmit the in-group social layer pre-aggregated model parameter updates to a global aggregator to obtain a global aggregated model, and the global aggregator transmits the global aggregated model to each user participating in the federated learning task.

[0188] In one embodiment of the present invention, a computer device is provided, which includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function. The processor described in the embodiment of the present invention can be used to implement the operation of a federated learning method based on social grouping.

[0189] In one embodiment of the present invention, when a federated learning method based on social grouping is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable storage medium includes permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data.

[0190] The computer storage medium can be any available medium or data storage device accessible by a computer, including but not limited to magnetic memories (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MO), etc.), optical memories (such as CDs, DVDs, BDs, HVDs, etc.), and semiconductor memories (such as ROMs, EPROMs, EEPROMs, non-volatile memories (NAND FLASH), solid state drives (SSD)).

[0191] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0192] The present application is described with reference to the flowcharts and / or block diagrams of the methods, apparatuses (systems), and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the specified functions in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0193] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the specified functions in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0194] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the specified functions in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0195] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any technician familiar with the technical field of the present invention can still modify the technical solutions described in the foregoing embodiments or can easily think of changes, or perform equivalent replacements for some of the technical features; and these modifications, changes or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims described herein.

Claims

1. A federated learning method based on social grouping, characterized in that, Including: For users jointly participating in the federated learning task, construct a social relationship graph among users. In the social relationship graph, each user evaluates the social trust value of other users based on direct social trust and indirect social trust; The users jointly participating in the federated learning task form multiple social groups that are Nash stable and non - overlapping according to the social trust value. Specifically: For each social group, define a group utility function according to the social trust value among users; For each social group, design a fair distribution rule according to the group utility function to obtain the individual utility functions of all members within the social group; For each user participating in the federated learning task, construct a transferable group list according to the individual utility function, and send a join application to the optimal social group in the transferable group list while satisfying the social group departure rule; For each social group, when receiving multiple join applications, construct an acceptable user list, and accept the join application of the optimal user in the acceptable user list while satisfying the social group join rule, and reject the join applications of other users in the acceptable user list; Execute user separation and merger operations, and update the set of current social groups; Repeat the above steps until the state of the social groups no longer changes, forming the multiple social groups that are Nash stable and non - overlapping; All members within each social group use local data for training to obtain updated local model parameters, and determine a Gaussian noise perturbation strategy according to the social trust value with the manager of the social group to obtain the perturbed updated local model parameters. The manager of the social group is the user with the highest centrality within the social group; The manager of each social group performs pre - aggregation of the models within the social group on the updated local model parameters of all members within the social group to obtain updated social - layer pre - aggregated model parameters; Each social group transmits the updated social - layer pre - aggregated model parameters to the global aggregator to obtain a global aggregated model, and the global aggregator transmits the global aggregated model to each user participating in the federated learning task.

2. The federated learning method based on social grouping according to claim 1, wherein The specific method for defining the group utility function according to the social trust value among users is: The social grouping The group utility is the difference between the benefit and the cost: Wherein, is the reward payment for the quality of the unit model; is the social grouping in which the user n the quality of the local model, is related to the social relationship between the user n and the social grouping manager and the degree of non-independent and identically distributed of the user local data; n is related to the degree of non-independent and identically distributed of the local data of the user is a proportionality coefficient greater than 0; represents the size of the social grouping i.e., the number of members within the social grouping; represents the communication cost within the social grouping ; represents the model quality when the user n forms a single-item grouping; The social grouping in which the user n local model quality and the model loss value are negatively correlated, specifically: wherein, are all positive curve fitting parameters; represents the asymptotic maximum local model quality when the model loss tends to 0; The social grouping The user n Local model loss Is related to the Gaussian noise magnitude And the degree of non-independent and identically distributed data Specifically: In the formula, are all positive curve fitting parameters; the numerator part reflects the decrease in the marginal loss value of the local model when the degree of non-independent and identically distributed data increases; the denominator part reflects the magnitude of the model noise and the decline in the quality of the local model when it increases.

3. The federated learning method based on social grouping according to claim 2, wherein, The specific method for designing a fair distribution rule according to the group utility function to obtain the individual utility functions of all members within the social group is: The social grouping Each member user n The individual utility is the sum of the utility in the non - cooperative case and the additional utility distributed according to the contribution ratio: Wherein, and respectively represent the non - cooperative utilities of users l and n , that is, the group utilities when users l and n form single - item groups respectively; represents the contribution weight of user n in the social group ; represents the additional utility generated when all members in the social group cooperate compared with the non - cooperative situation.

4. A federated learning method based on social grouping according to claim 3, characterized in that The specific method for constructing a transferable group list according to the individual utility function is: Construct a preference order according to the individual utility function; Construct a transferable group list according to the preference order; The specific method for constructing a preference order according to the individual utility function is: User The preference order for social groups is as follows: In the formula, represents the individual utility of the user n after joining the social group ; and respectively represent the individual utilities of the user in the original social group and the new social group ; represents the set of social groups that have historically rejected the user's n application to join; The specific method for constructing a transferable group list according to the preference order is: At the t th iteration, the list of transferable groups of users in the social group is as follows: n ​ In the formula, represents the set of all social groups t at the t -th iteration.

5. The federated learning method based on social grouping according to claim 4, characterized in that, The social group departure rule is specifically: If the social group accepts a new user's join application at the t th iteration, then none of the original internal members can leave the social group at the t th iteration ; The optimal social group in the transferable group list is specifically: User The optimal social group in the transferable group list of ; If , then the user n tends to break away from the original social group and form a single-item group; otherwise, the user n tends to break away from the original social group and join the social group .

6. The federated learning method based on social grouping according to claim 5, characterized in that The social group join rule is specifically: If an internal member leaves the social group during the t th iteration, the social group cannot accept the joining applications of all new users during the t th iteration; The optimal user in the acceptable user list is specifically: Social grouping The optimal user in the acceptable user list of is where t is the set of acceptable users of the social grouping at the th iteration, that is, the set of users who send join requests to the social grouping t at the th iteration.

7. A federated learning method based on social grouping according to claim 1, characterized in that, The specific method for determining the Gaussian noise perturbation strategy according to the social trust value with the manager of the social group is: If the user n has a social trust value greater than the system - preset threshold with the manager of the social group then directly update and transmit the original local model parameters to the manager of the social group ; If the user n and the social trust value between the user and the manager of the social group is lower than the system preset threshold, then the updated local model parameters with Gaussian noise perturbation are transmitted to the manager of the social group ; If the user n forms a single-item group, a maximum Gaussian noise perturbation of size is added to the local model parameter update, and the user n is the manager of the single-item group.

8. A federated learning method based on social grouping according to claim 7, characterized in that, The update of the local model parameters with Gaussian noise perturbation is specifically as follows: Each member within each social group maps the social trust value between the member and the administrator of the social group to a privacy protection level; Each member within each social group calculates the magnitude parameter of Gaussian noise according to the privacy protection level; Each member within each social group adds corresponding Gaussian noise to the update of the local model parameters according to the magnitude parameter of the Gaussian noise, and obtains the updated local model parameters after perturbation; The mapping of the social trust value between the member and the administrator of the social group to a privacy protection level is specifically as follows: In the formula, is a positive adjustment parameter; represents the privacy budget, and the smaller its value, the greater the degree of privacy protection; represents the user n and the social group the social trust value between the managers; Each member within each social group calculates the magnitude parameter of Gaussian noise according to the privacy protection level, specifically as follows: wherein, is a small probability value; is the variance value of Gaussian noise, and the larger its value, the greater the magnitude of the added noise; , represents the maximum value of the Gaussian noise variance.

9. A federated learning device based on social grouping, characterized in that, It includes: A construction module, which is used to construct a social relationship graph among users for users jointly participating in the federated learning task. In the social relationship graph, each user evaluates the social trust value of other users based on direct social trust and indirect social trust; A social group formation module, which is used for the users jointly participating in the federated learning task to form multiple social groups that are Nash stable and non - overlapping according to the social trust value. Specifically: For each social group, define a group utility function according to the social trust value between users; For each social group, design a fair distribution rule according to the group utility function to obtain the individual utility functions of all members within the social group; For each user participating in the federated learning task, construct a transferable group list according to the individual utility function, and send a join application to the optimal social group in the transferable group list while satisfying the social group leaving rule; For each social group, when receiving multiple join applications, construct an acceptable user list, and accept the join application of the optimal user in the acceptable user list while satisfying the social group joining rule, and reject the join applications of other users in the acceptable user list at the same time; Execute user separation and merger operations, and update the set of current social groups; Repeat the above steps until the state of the social groups no longer changes, and form the multiple social groups that are Nash stable and non - overlapping; A local model parameter update module, which is used for all members within each social group to train using local data to obtain the update of the local model parameters, and determine the Gaussian noise perturbation strategy according to the social trust value between the member and the administrator of the social group, and obtain the updated local model parameters after perturbation. The administrator of the social group is the user with the highest centrality within the social group; A pre - aggregation module, which is used for the administrator of each social group to perform model pre - aggregation within the social group on the updates of the local model parameters of all members within the social group to obtain the updated social layer pre - aggregation model parameters; A transmission module, which is used for each social group to transmit the updated social layer pre - aggregation model parameters to the global aggregator to obtain the global aggregation model, and the global aggregator transmits the global aggregation model to each user participating in the federated learning task.

Images