A model parameter updating method and system based on graph federated learning

Abstract

The application discloses a model parameter updating method and system based on graph federated learning, relates to the fields of federated learning and graph neural networks, and uses the framework of federated learning to complete training of a graph convolutional neural network, so that the graph convolutional neural network can train a model on graphs distributed in multiple clients; a neighbor feature aggregation function is used, so that the clients do not need to communicate in each training iteration, and the communication burden is significantly reduced; an aggregation mechanism based on an attention mechanism is used, the server is allowed to dynamically allocate weights for each local model participating in training, and the robustness and accuracy of the model are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mobile communication, and particularly relates to a model parameter updating method and system based on graph federated learning. BACKGROUND

[0002] Currently, in order to protect the security of participant data privacy in the joint training model, the federated learning technology emerges as the times require. Federated learning is a promising distributed collaborative machine learning paradigm, aiming to solve the problem of data isolation. It allows local models to promote each other while ensuring the privacy of local data. In addition, personalized federated learning models under non-independent and identically distributed conditions have also attracted widespread attention. In federated learning, participants only share model parameters or embedding features, avoiding the leakage of sensitive data. This architecture combines various encryption technologies, effectively protecting the privacy of local data. At the same time, there may be relationships between federated learning participants, and the similarity and trust information of the participants can be revealed through the constructed graph topology. It still faces challenges to improve the performance of federated learning by using these topological information. Therefore, combining federated learning with graph machine learning is a very promising solution.

[0003] Currently, the challenges of graph federated learning are as follows: in the environment of federated learning, data is divided and stored on multiple client participants in the training. The interaction and connection between these data is usually embodied through cross-client edges, which play a crucial role in graph neural networks. They not only connect the graph nodes of data stored by different clients, but also capture the complex relationships and dependencies between graph nodes, which are crucial for the accuracy and robustness of the model. However, traditional federated learning methods often do not fully consider the existence and importance of these cross-client edges, resulting in the model being unable to fully capture global structural information when using graph data, thereby affecting the performance of the model.

[0004] At the same time, federated learning systems face special challenges when dealing with non-Euclidean space data, which often contain complex graph structures such as social networks and biological information networks. In order to improve the performance of federated learning on such data, researchers have been exploring more advanced aggregation strategies. However, traditional federated learning methods often use simple average aggregation strategies, which do not fully consider the uniqueness and quality differences of each local model, and may lead to a decrease in the performance of the aggregated global model. SUMMARY

[0005] In order to solve the problems of cross-client edges being easily ignored and the poor aggregation effect of traditional federated learning methods, the application proposes a model parameter updating method and system based on graph federated learning.

[0006] In a first aspect, the present application provides a model parameter updating method based on graph federated learning, applied to a server of graph federated learning, comprising the following steps:

[0007] initializing global model parameters and broadcasting to each client;

[0008] receiving local model parameters trained by the client;

[0009] evaluating the local model parameters of each client to obtain the first confidence of the local model parameters of each client; the first confidence of the local model parameters of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client;

[0010] quantum state encoding the local model parameters of each client to obtain the quantum state model parameters of each client;

[0011] operating the quantum state model parameters of each client through a quantum self-attention circuit to obtain the second confidence of the local model parameters of each client; the second confidence of the local model parameters of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients;

[0012] updating the global model parameters of the server according to the first confidence and the second confidence of the local model parameters of each client;

[0013] broadcasting the updated global model parameters to all clients;

[0014] if the global model parameters converge, ending the updating, otherwise repeating the above process.

[0015] In a second aspect, the present application provides a model parameter updating method based on graph federated learning, applied to a client of graph federated learning, comprising the following steps:

[0016] receiving global model parameters from the server; the global model parameters are updated based on the first confidence and the second confidence of the local model parameters of each client; the first confidence of the local model parameters of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client; the second confidence of the local model parameters of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients, which is obtained by operating the quantum state model parameters of each client through a quantum self-attention circuit; the quantum state model parameters of each client are obtained by quantum state encoding the local model parameters of each client;

[0017] constructing a graph convolutional neural network according to the global model parameter and the multi-hop neighbor aggregated feature;

[0018] training the global model parameter to obtain an updated local model parameter;

[0019] sending the updated local model parameter to the server;

[0020] ending the updating until the global model parameter from the server is not received within a preset time, otherwise repeating the above process; the preset time is used to indicate a convergence time of the global model parameter.

[0021] In a third aspect of the present application, the present application provides a model parameter updating system based on graph federated learning, deployed in a server participating in graph federated learning, the model parameter updating system based on graph federated learning comprising:

[0022] a processor; and

[0023] a memory arranged to store computer executable instructions which, when executed, cause the processor to:

[0024] initializing a global model parameter and broadcasting to each client;

[0025] receiving a local model parameter trained from the client;

[0026] evaluating the local model parameter of each client to obtain a first confidence of the local model parameter of each client; the first confidence of the local model parameter of each client is determined by the similarity between the global model parameter of the server and the local model parameter of each client;

[0027] quantum state encoding the local model parameter of each client to obtain a quantum state model parameter of each client;

[0028] operating the quantum state model parameter of each client through a quantum self-attention circuit to obtain a second confidence of the local model parameter of each client; the second confidence of the local model parameter of each client is determined by the similarity between the local model parameter of each client and the local model parameter of other clients;

[0029] updating the global model parameter of the server according to the first confidence and the second confidence of the local model parameter of each client;

[0030] broadcasting the updated global model parameter to all clients;

[0031] ending the updating if the global model parameter converges, otherwise repeating the above process.

[0032] In a fourth aspect of the present application, the present application provides a model parameter updating system based on graph federated learning, deployed on a client participating in the graph federated learning, comprising:

[0033] a processor; and

[0034] a memory arranged to store computer executable instructions that, when executed, cause the processor to:

[0035] receive global model parameters from a server; the global model parameters are updated based on a first confidence and a second confidence of a local model parameter of each client; the first confidence of the local model parameter of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client; the second confidence of the local model parameter of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients, which is obtained by quantum self-attention circuit operation of quantum state model parameters of each client; the quantum state model parameters of each client are obtained by quantum state encoding of the local model parameters of each client;

[0036] construct a graph convolutional neural network according to the global model parameters and multi-hop neighbor aggregated features;

[0037] train the global model parameters to obtain updated local model parameters;

[0038] send the updated local model parameters to the server;

[0039] until the global model parameters from the server are not received within a preset time, the updating is ended, otherwise the above process is repeated; the preset time is used to indicate the convergence time of the global model parameters.

[0040] The present application has the following beneficial effects:

[0041] In the federated graph convolutional neural network, in order to maintain the accuracy of the model, it is necessary to exchange the information of the node neighbors in the training process, which leads to significant communication overhead. The present application significantly reduces the communication cost by communicating only a limited number of times in the pre-training step. The present application also uses an attention-based aggregation mechanism that can dynamically assign weights to each local model participating in training. This attention-based approach not only takes into account the individual differences of local models, but also adjusts their influence in the aggregation process according to their contribution to the global model. In this way, the present application can identify and reduce the negative impact of local models with poor performance or uneven data distribution on the global model. This not only improves the fault tolerance of the model when facing clients with varying data quality, but also enhances the accuracy of the model under various data distributions. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a model parameter updating method flowchart based on graph federated learning according to an embodiment of the present application;

[0043] Figure 2 is a model parameter updating method flowchart in the pre-training stage according to an embodiment of the present application;

[0044] Figure 3 is a model parameter updating method flowchart based on graph federated learning according to another embodiment of the present application;

[0045] Figure 4 is a model parameter updating method flowchart in the pre-training stage according to another embodiment of the present application;

[0046] Figure 5 is a signaling diagram of a model parameter updating method based on graph federated learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] The embodiments of the present application provide a model parameter updating method based on graph federated learning. The execution subject of the method can be a participating device of graph federated learning. The participating device of graph federated learning can be a client or a server. The server can be an independent server or a server cluster composed of multiple servers. The client can be a personal computer or a terminal device with display function such as a smartphone, a smart television, a tablet computer, and a portable computer.

[0049] Figure 1 is a flow chart of a model parameter updating method based on graph federated learning according to an embodiment of the present application, as shown in Figure 1 the method comprises the following steps:

[0050] 101. Initialize the global model parameters and broadcast them to each client;

[0051] In some embodiments, when starting training, the server initializes the global model parameters and broadcasts them to all clients under graph federated learning, where superscript t represents the iteration round, t=0 initially, and subscript g is used to indicate the global model of the server.

[0052] 102. Receive the local model parameters trained by the clients;

[0053] In some embodiments, after entering training, the server will receive the local model parameters w k returned by the clients, where subscript k is used to indicate the local model of the kth client.

[0054] 103. Evaluate the local model parameters of each client to obtain the first confidence of the local model parameters of each client; the first confidence of the local model parameters of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client;

[0055] In some embodiments, the evaluation of the local model parameters of each client to obtain the first confidence of the local model parameters of each client comprises:

[0056] Similarity measurement is performed between the local model parameters of each client and the global model parameters of the server to obtain the similarity distance between the local model parameters of each client and the global model parameters of the server;

[0057] The similarity distance between the local model parameters of each client and the global model parameters of the server is subjected to attention processing by a standardization function to obtain the first confidence of the local model parameters of each client.

[0058] For example, assume is the trainable parameter of the local model w k of the kth client at the lth layer, is the trainable parameter of the global model w g at the lth layer, and the similarity distance between the local model parameters of the kth client and the global model parameters of the server at the lth layer is obtained by ​The attention coefficient of the local model parameter of the kth client at the lth layer to the global model parameter of the server can be further characterized; in order to facilitate the use of the coefficients in the layer parameters, and not to affect the convergence speed due to the possible order of magnitude transformation of the parameters, the SoftMax function is used to normalize the attention coefficient, thereby obtaining the first confidence of the local model parameter of the kth client wherein the superscript L represents the total number of layers, and the subscript k represents the kth client.

[0059] 104. Quantum state encoding is performed on the local model parameters of each client to obtain quantum state model parameters of each client.

[0060] In some embodiments, the server normalizes the local model parameters w k Encoding into quantum states Specifically represented as:

[0061]

[0062] wherein H represents a Hadamard gate, denotes a tensor operator, U enc denotes quantum amplitude encoding, and n represents the number of clients.

[0063] 105. Quantum self-attention circuit is used to operate on the quantum state model parameters of each client to obtain the second confidence of the local model parameters of each client; the second confidence of the local model parameters of each client is determined by the similarity of the local model parameters of each client and the local model parameters of other clients.

[0064] In some embodiments, the calculation process of the similarity of the local model parameters between each client includes:

[0065] The quantum state model parameters of all clients are respectively subjected to corresponding parameter evolution, and Z-direction measurement is performed according to the Pauli-Z gate to obtain the first measurement expectation and the second measurement expectation corresponding to the quantum state model parameters of each client.

[0066] According to the difference between the first measurement expectation and the second measurement expectation corresponding to the quantum state model parameters of each client, the quantum attention matrix of the corresponding client is obtained.

[0067] According to the quantum attention matrix of each client, the similarity of the local model parameters between each client and other clients is calculated.

[0068] According to the similarity of the local model parameters between each client and other clients, the second confidence of the local model parameters of each client is calculated.

[0069] In some embodiments, the server obtains the quantum states of all clients In some embodiments, the server obtains the quantum states of all clients In some embodiments, the server obtains the quantum states of all clients

[0070] For example, the encoded quantum states are evolved through the corresponding parameter circuits, and finally measured in the Z direction according to the Pauli-Z gate to obtain the measurement expectation, which is specifically represented as:

[0071]

[0072]

[0073] wherein U K and U Q are quantum circuits, U K and U Q are the same except for the parameters θ K and θ Q contained therein. <Z K > and <Z Q > represent the measurement expectation obtained by measuring in the Z direction according to the Pauli-Z gate. denotes the conjugate transpose of a matrix. Z represents the Pauli-Z gate.

[0074] Based on the first measurement expectation and the second measurement expectation obtained according to the above calculation formula, the corresponding attention matrix is obtained, which is specifically represented as:

[0075]

[0076] According to the obtained attention matrix, further calculation is performed to obtain the attention similarity of any l of each client k, and then the original similarity value is mapped to a probability distribution with a sum of 1 through the softmax function, which is specifically represented as:

[0077]

[0078] It can be understood that the calculation cost of the traditional self-attention mechanism increases sharply with the increase of the number of parameters, while the self-attention mechanism implemented by the quantum circuit can more effectively manage the calculation demand of a large number of parameters and more efficiently cope with the calculation challenge of large-scale parameters by means of the parallel processing capability of quantum computing, thereby optimizing the use of computing resources. The embodiments of the present application can dynamically allocate weights to each local model participating in training. This self-attention-based method not only considers the individual differences of local models, but also adjusts their influence in the aggregation process according to their contribution to the global model. In this way, the negative impact of local models with poor performance or uneven data distribution on the global model can be identified and reduced. This not only improves the fault tolerance of the model when facing clients with uneven data quality, but also enhances the accuracy of the model under various data distributions.

[0079] 106、According to the first confidence and the second confidence of the local model parameters of each client, updating the global model parameters of the server;

[0080] In some embodiments, the server updates the global model according to the first confidence determined by the similarity of the model parameters between the client and the server, and the second confidence determined by the similarity of the model parameters between the clients.

[0081] For example, the updating rule of the global model is specifically expressed as:

[0082]

[0083] wherein p and q are adjustable hyperparameters, represents the first confidence of the local model parameters of the kth client at the lth layer in the tth iteration process, represents the second confidence of the local model parameters of the kth client at the lth layer in the tth iteration process; by fusing the information of the server and other clients through the two confidences, the local model of each client participating in training can be dynamically allocated weights, not only considering the individual differences of each local model, but also adjusting its influence in the aggregation process according to its contribution to the global model. In this way, the negative impact of local models with poor performance or uneven data distribution on the global model can be identified and reduced. This not only improves the fault tolerance of the model when facing clients with uneven data quality, but also enhances the accuracy of the model under various data distributions.

[0084] 107、Broadcasting the updated global model parameters to all clients;

[0085] In some embodiments, the updated global model parameters are broadcast to all clients, and the global model parameters can be updated by transitioning to complete a round of iterative training.

[0086] 108、if the global model parameters converge, end the update, otherwise repeat the above process.

[0087] In some embodiments, it is determined whether the global model converges, if not, steps 101-107 are repeated multiple times, if the global model converges, the training is ended and the updating of the model parameters is stopped.

[0088] In some preferred embodiments, as shown in Figure 2 Before the initialization of the global model parameters and the broadcasting to each client, the following steps are further included:

[0089] 111、receiving neighbor feature ciphertexts of the local graph node from the client;

[0090] 121、aggregating the neighbor feature ciphertexts of the local graph node of all clients to obtain multi-hop neighbor feature ciphertexts of the local graph node;

[0091] In some embodiments, after the server receives all the neighbor feature ciphertexts of the local graph node from the client, the neighbor aggregation operation is performed, which can be represented as:

[0092]

[0093] wherein c(j) represents the index of the client containing the graph node j, is an indicator function, when z is equal to c(j), that is, the client z contains the information of the graph node j, the indicator function has a value of 1; otherwise, the value is 0. is an encryption operator, represents the neighbor node of the graph node i, and the server accumulates the multi-hop neighbor features of each graph node by operating on all the received neighbor feature arrays.

[0094] 131、sending the multi-hop neighbor feature ciphertexts of the local graph node to the client.

[0095] This embodiment takes into account that in the federated graph convolutional neural network, in order to maintain the accuracy of the model, the information of the node neighbors needs to be exchanged constantly during the training process, which results in significant communication overhead. This embodiment adds the pre-training steps of steps 111-131 before the model parameter updating, only needs to perform a limited number of communications, significantly reduces the communication cost, and saves the communication overhead.

[0096] Figure 3 is a model parameter updating method flowchart based on graph federated learning according to an embodiment of the present application, as shown in Figure 3As shown, the method applied to a client of a federated learning of a graph comprises the following steps:

[0097] 201. receiving global model parameters from a server; the global model parameters are updated based on a first confidence and a second confidence of a local model parameter of each client; the first confidence of the local model parameter of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client; the second confidence of the local model parameter of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients, and the similarity between the local model parameters of each client and the local model parameters of other clients is obtained by quantum state model parameters of each client through a quantum self-attention circuit; the quantum state model parameters of each client are obtained by quantum state encoding of the local model parameters of each client;

[0098] In some embodiments, in a training round, the client continuously receives global model parameters issued by the server in an iterative round wherein the superscript t represents the iterative round, and the subscript g is used to indicate the global model of the server.

[0099] 202. constructing a graph convolutional neural network according to the global model parameters and multi-hop neighbor aggregated features;

[0100] In some embodiments, the construction of the graph convolutional neural network according to the global model parameters and multi-hop neighbor aggregated features comprises:

[0101] constructing a graph node embedding vector for each convolutional layer according to the multi-hop neighbor aggregated features;

[0102] constructing a graph convolutional neural network according to the graph node embedding vector of each convolutional layer and the global model parameters.

[0103] For example, a multi-layer graph convolutional neural network is composed of multiple convolutional layers, each convolutional layer constructs a node embedding by aggregating the features of adjacent nodes and follows a propagation rule:

[0104]

[0105] wherein W (l) is the weight of each layer of the convolutional neural network to be learned, A ij is the weighted adjacency matrix, φ is the activation function, is the embedding of each neighbor j of the previous layer. The softmax function is used as the activation function of the last layer, and the specific expression is:

[0106]

[0107] wherein, zi Let be the i-th component of the input, and e be the base of the natural logarithm. This allows the output to be interpreted as the probability of nodes in each class. ReLU is used as the activation function in the first few layers. Specifically:

[0108] f(a) = max(0,a)

[0109] Where 'a' is the input value, and 'f(a)' is the output of the activation function. The loss function used is the cross-entropy loss function, specifically expressed as:

[0110]

[0111] Where n k It represents the amount of data owned by the kth client. Let i be the i-th feature contained in the data of the k-th client. Let be the i-th label contained in the data of the k-th client, and these labels are encoded using one-hot representation. The GCN model is denoted as G, and the node features X are input into the model to obtain the output G(X).

[0112] 203. Train the global model parameters to obtain the updated local model parameters;

[0113] In some embodiments, after receiving the global model from the server, the client trains the global model using its own data. The client-side local training uses gradient descent to update the model, and the specific update rules for each round are described as follows:

[0114]

[0115] in This represents the weights of the L GCN layers in the e-th local training step of client k during the t-th round of global training. Serial, f k Let η represent the local loss function, and η be the learning rate.

[0116] 204. Send the updated local model parameters to the server;

[0117] In some embodiments, in each iteration t, the client trains its local model w τ times locally. k It is sent to the server.

[0118] 205. The update ends if no global model parameters are received from the server within a preset time; otherwise, the above process is repeated. The preset time is used to indicate the convergence time of the global model parameters.

[0119] In some embodiments, if the client does not receive the global model parameters issued by the server within a preset time, it indicates that the server considers that the global model parameters have reached convergence, and the client can end the model update.

[0120] In some preferred embodiments, as shown in Figure 4 the receiving of the global model parameters from the server further comprises:

[0121] 211, encrypting the neighbor features of the local graph node using a homomorphic encryption algorithm to generate neighbor feature ciphertext of the local graph node;

[0122] In some embodiments, each client sends the neighbor features of the local graph node to the server, which can be specifically expressed as:

[0123]

[0124] wherein c(j) represents the index of the client containing the graph node j, is an indicator function, when k is equal to c(j), that is, the client z contains the information of the graph node j, the indicator function has a value of 1; otherwise, the value is 0. is an encryption operator, A ij represents a weighted adjacency matrix, x j represents the feature vector of the graph node.

[0125] 212, sending the neighbor feature ciphertext of the local graph node to the server of the graph federated learning;

[0126] 213, receiving the multi-hop neighbor feature ciphertext of the local graph node from the server;

[0127] In some embodiments, after the server receives all the neighbor feature ciphertexts of the local graph nodes from the clients, it performs a neighbor aggregation operation, which can be expressed as:

[0128]

[0129] wherein c(j) represents the index of the client containing the graph node j, is an indicator function, when z is equal to c(j), that is, the client z contains the information of the graph node j, the indicator function has a value of 1; otherwise, the value is 0. is an encryption operator, represents the neighbor node of the graph node i, A ij represents a weighted adjacency matrix, x j represents the feature vector of the graph node j, and the server accumulates the multi-hop neighbor features of each node by operating on all the received neighbor feature arrays.

[0130] 214. decrypting the multi-hop neighbor feature ciphertext of the local graph node to generate a multi-hop neighbor aggregated feature of the local graph node.

[0131] The embodiment considers that in the federated graph convolutional neural network, in order to maintain the accuracy of the model, the information of the node neighbors needs to be exchanged constantly in the training process, which causes significant communication overhead. The embodiment significantly reduces the communication cost and saves the communication overhead by adding the pre-training steps 211-214 before the model parameter update, only needing to perform a limited number of communications.

[0132] Figure 5 is a model parameter update scene diagram based on graph federated learning according to an embodiment of the application, as shown in Figure 5 The method comprises the following steps of:

[0133] The client uses a homomorphic encryption algorithm to encrypt the neighbor features of the local graph node to generate neighbor feature ciphertext of the local graph node;

[0134] The client sends the neighbor feature ciphertext of the local graph node to the server of the graph federated learning;

[0135] The server aggregates the neighbor feature ciphertext of the local graph node of all clients to obtain multi-hop neighbor feature ciphertext of the local graph node;

[0136] The server sends the multi-hop neighbor feature ciphertext of the local graph node to the client;

[0137] The client decrypts the multi-hop neighbor feature ciphertext of the local graph node to generate a multi-hop neighbor aggregated feature of the local graph node;

[0138] The server initializes the global model parameters and broadcasts them to each client;

[0139] The client constructs a graph convolutional neural network according to the global model parameters and the multi-hop neighbor aggregated feature;

[0140] The client trains the global model parameters to obtain updated local model parameters;

[0141] The server evaluates the local model parameters of each client to obtain a first confidence of the local model parameters of each client;

[0142] The server quantum state encodes the local model parameters of each client to obtain quantum state model parameters of each client;

[0143] The server operates the quantum state model parameters of each client through a quantum self-attention circuit to obtain a second confidence of the local model parameters of each client.

[0144] The server updates the global model parameters of the server according to the first confidence and the second confidence of the local model parameters of each client;

[0145] The server broadcasts the updated global model parameters to all clients;

[0146] If the global model parameters converge, the updating is ended, otherwise the above process is repeated.

[0147] A model parameter updating system based on graph federated learning according to an embodiment of the present application is deployed on a server participating in graph federated learning, and the model parameter updating system based on graph federated learning comprises:

[0148] a processor; and

[0149] a memory arranged to store computer executable instructions that, when executed, cause the processor to:

[0150] initialize the global model parameters and broadcast them to each client;

[0151] receive the local model parameters trained by each client;

[0152] evaluate the local model parameters of each client to obtain the first confidence of the local model parameters of each client; the first confidence of the local model parameters of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client;

[0153] quantum state encode the local model parameters of each client to obtain the quantum state model parameters of each client;

[0154] operate the quantum state model parameters of each client through a quantum self-attention circuit to obtain the second confidence of the local model parameters of each client; the second confidence of the local model parameters of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients;

[0155] update the global model parameters of the server according to the first confidence and the second confidence of the local model parameters of each client;

[0156] broadcast the updated global model parameters to all clients;

[0157] If the global model parameters converge, the updating is ended, otherwise the above process is repeated.

[0158] A model parameter updating system based on graph federated learning is deployed on a client participating in graph federated learning, and comprises:

[0159] a processor; and

[0160] a memory arranged to store computer executable instructions that, when executed, cause the processor to:

[0161] receive global model parameters from a server; the global model parameters are updated based on a first confidence and a second confidence of a local model parameter of each client; the first confidence of the local model parameter of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client; the second confidence of the local model parameter of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients, and the similarity between the local model parameters of each client and the local model parameters of other clients is obtained by quantum state model parameters of each client through quantum self-attention circuit; the quantum state model parameters of each client are obtained by quantum state encoding of the local model parameters of each client;

[0162] construct a graph convolutional neural network according to the global model parameters and multi-hop neighbor aggregated features;

[0163] train the global model parameters to obtain updated local model parameters;

[0164] send the updated local model parameters to the server;

[0165] until the global model parameters from the server are not received within a preset time, the updating is ended, otherwise the above process is repeated; the preset time is used to indicate the convergence time of the global model parameters.

[0166] The system, device, module or unit illustrated in the above embodiments can be specifically implemented by a computer chip or entity, or by a product with certain functions. A typical implementation device is a computer. Specifically, the computer may, for example, be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0167] The embodiments of the present specification are described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present specification. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and a combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one flow or a plurality of flows and / or blocks Figure 1 one flow or a plurality of flows and / or blocks

[0168] A person of ordinary skill in the art can understand that all or part of the steps in the above-mentioned embodiments can be completed by instructing the relevant hardware by a program, and the program can be stored in a computer readable storage medium, which can include ROM, RAM, magnetic disk or optical disk, etc.

[0169] Although the embodiments of the present application have been shown and described, it should be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A model parameter update method based on graph federated learning, characterized in that, A server applied to graph federation learning, the method comprising the following steps: Initialize the global model parameters and broadcast them to all clients; Receive local model parameters from the client after training; The local model parameters of each client are evaluated to obtain the first confidence level of the local model parameters of each client; the first confidence level of the local model parameters of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client. Quantum state encoding is performed on the local model parameters of each client to obtain the quantum state model parameters of each client; The quantum state model parameters of each client are calculated through a quantum self-attention circuit to obtain the second confidence level of the local model parameters of each client; the second confidence level of the local model parameters of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients. Update the server's global model parameters based on the first and second confidence scores of each client's local model parameters; Broadcast the updated global model parameters to all clients; If the global model parameters converge, the update ends; otherwise, the above process is repeated. Before initializing the global model parameters and broadcasting them to each client, the following steps are also included: Receive ciphertext of neighbor features from local graph nodes of the client; Aggregate the ciphertext of neighbor features of local graph nodes of all clients to obtain the ciphertext of multi-hop neighbor features of local graph nodes; The encrypted multi-hop neighbor features of the local graph node are sent to the client.

2. The model parameter update method based on graph federated learning according to claim 1, characterized in that, The evaluation of the local model parameters for each client to obtain the first confidence level of the local model parameters for each client includes: The similarity between the local model parameters of each client and the global model parameters of the server is measured to obtain the similarity distance between the local model parameters of each client and the global model parameters of the server. Attention processing is performed using a normalization function to measure the similarity distance between the local model parameters of each client and the global model parameters of the server, thereby obtaining the first confidence score of the local model parameters of each client.

3. The model parameter update method based on graph federated learning according to claim 1, characterized in that, The second confidence level of the local model parameters for each client is obtained by calculating the quantum state model parameters for each client through a quantum self-attention circuit, including: The quantum state model parameters of all clients are evolved according to their corresponding parameters, and the Z-direction is measured according to the Pauli-Z gate to obtain the first measurement expectation and the second measurement expectation corresponding to the quantum state model parameters of each client. The quantum attention matrix of the corresponding client is obtained based on the difference between the first and second measurement expectations corresponding to the quantum state model parameters of each client. Based on the quantum attention matrix of each client, the similarity of local model parameters between each client and other clients is calculated; The second confidence level of each client's local model parameters is calculated based on the similarity of the local model parameters between each client and other clients.

4. A model parameter update method based on graph federated learning, characterized in that, The method, applied to a client for graph federation learning, includes the following steps: The system receives global model parameters from the server; these global model parameters are updated based on a first confidence level and a second confidence level of each client's local model parameters; the first confidence level of each client's local model parameters is determined by the similarity between the server's global model parameters and each client's local model parameters; the second confidence level of each client's local model parameters is determined by the similarity between each client's local model parameters and other clients' local model parameters, and the similarity between each client's local model parameters and other clients' local model parameters is obtained by operating on each client's quantum state model parameters through a quantum self-attention circuit; the quantum state model parameters of each client are obtained by quantum state encoding of each client's local model parameters. Construct a graph convolutional neural network based on the global model parameters and multi-hop neighbor aggregation features; The global model parameters are trained to obtain the updated local model parameters; Send the updated local model parameters to the server; The update process ends if no global model parameters are received from the server within a preset time; otherwise, the process is repeated. The preset time is used to indicate the convergence time of the global model parameters.

5. The model parameter update method based on graph federated learning according to claim 4, characterized in that, The steps preceding receiving the global model parameters from the server are as follows: The neighbor features of local graph nodes are encrypted using a homomorphic encryption algorithm to generate ciphertext of the neighbor features of local graph nodes. Send the encrypted neighbor features of the local graph node to the server for graph federation learning; Receive multi-hop neighbor feature ciphertext from local graph nodes on the server; The ciphertext of the multi-hop neighbor features of the local graph node is decrypted to generate the multi-hop neighbor aggregation feature.

6. A method for updating model parameters based on graph federated learning according to claim 4 or 5, characterized in that, The step of constructing a graph convolutional neural network based on the global model parameters and multi-hop neighbor aggregation features includes: Construct graph node embedding vectors for each convolutional layer based on multi-hop neighbor aggregation features; A graph convolutional neural network is constructed based on the graph node embedding vectors of each convolutional layer and the global model parameters.

7. A model parameter update system based on graph federated learning, characterized in that, Deployed on servers participating in graph federated learning, the graph federated learning-based model parameter update system includes: Processor; and A memory configured to store computer-executable instructions, which, when executed, cause the processor to: Initialize the global model parameters and broadcast them to all clients; Receive local model parameters from the client after training; The local model parameters of each client are evaluated to obtain the first confidence level of the local model parameters of each client; the first confidence level of the local model parameters of each client is determined by the similarity between the global model parameters of the server and the local model parameters of each client. Quantum state encoding is performed on the local model parameters of each client to obtain the quantum state model parameters of each client; The quantum state model parameters of each client are calculated through a quantum self-attention circuit to obtain the second confidence level of the local model parameters of each client; the second confidence level of the local model parameters of each client is determined by the similarity between the local model parameters of each client and the local model parameters of other clients. Update the server's global model parameters based on the first and second confidence scores of each client's local model parameters; Broadcast the updated global model parameters to all clients; If the global model parameters converge, the update ends; otherwise, the above process is repeated. Before initializing the global model parameters and broadcasting them to each client, the following steps are also included: Receive ciphertext of neighbor features from local graph nodes of the client; Aggregate the ciphertext of neighbor features of local graph nodes of all clients to obtain the ciphertext of multi-hop neighbor features of local graph nodes; The encrypted multi-hop neighbor features of the local graph node are sent to the client.

8. A model parameter update system based on graph federated learning, characterized in that, Deployed on clients participating in graph federated learning, the graph federated learning-based model parameter update system includes: Processor; and A memory configured to store computer-executable instructions, which, when executed, cause the processor to: The system receives global model parameters from the server; these global model parameters are updated based on a first confidence level and a second confidence level of each client's local model parameters; the first confidence level of each client's local model parameters is determined by the similarity between the server's global model parameters and each client's local model parameters; the second confidence level of each client's local model parameters is determined by the similarity between each client's local model parameters and other clients' local model parameters, and the similarity between each client's local model parameters and other clients' local model parameters is obtained by operating on each client's quantum state model parameters through a quantum self-attention circuit; the quantum state model parameters of each client are obtained by quantum state encoding of each client's local model parameters. Construct a graph convolutional neural network based on the global model parameters and multi-hop neighbor aggregation features; The global model parameters are trained to obtain the updated local model parameters; Send the updated local model parameters to the server; The update process ends if no global model parameters are received from the server within a preset time; otherwise, the process is repeated. The preset time is used to indicate the convergence time of the global model parameters.

Images