A communication optimization method and system for quantum federated learning

Abstract

This invention relates to a communication optimization method and system for quantum federated learning. The method includes: based on a constructed communication optimization objective, each client receives a global model from the server and trains it for the current round; after completing the current round of training, each client quantizes the format of the trained global model parameters and uploads them to the server; the server receives the quantized global model parameters, constructs the final aggregation weights corresponding to the global model parameters of each client, and aggregates the global model parameters of each client according to the aggregation weights to obtain an updated global model; the server dynamically prunes the updated global model and quantizes the format of the retained global model parameters to generate a compressed model, which is then distributed to each client; the above process is repeated until a preset number of training rounds are completed. This invention can effectively reduce the communication overhead of quantum federated learning.

Description

Technical Field

[0001] This invention relates to the field of communication optimization technology in quantum federated learning, and in particular to a communication optimization method and system for quantum federated learning. Background Technology

[0002] Federated Learning (FL), as an emerging distributed machine learning paradigm, allows multiple data holders to collaboratively train models without sharing the original data, effectively solving the problems of data privacy and data silos, and showing great application potential in fields such as healthcare, finance, and the Internet of Things. However, as model complexity increases, federated learning faces severe communication bottlenecks, with frequent model parameter synchronization placing a heavy burden on network bandwidth.

[0003] Meanwhile, the rapid development of quantum computing (QC), especially the emergence of noisy intermediate-scale quantum (NISQ) devices, has provided a new computational paradigm for machine learning. Quantum machine learning (QML) combines the parallelism of quantum computing with classical machine learning algorithms, utilizing models such as parameterized quantum circuits (PQC), and has shown the potential to surpass classical models on specific tasks. Among them, quantum convolutional neural networks (QCNNs), as a structured quantum model, have become a powerful tool for processing complex data due to their parameter sharing and hierarchical abstraction characteristics.

[0004] The convergence of these two technological trends has given rise to Quantum Federated Learning (QFL). QFL aims to jointly train a global quantum model using distributed quantum or hybrid classical-quantum devices, thereby achieving both the high-dimensional representation advantages of quantum computing and the privacy protection mechanisms of federated learning.

[0005] Existing quantum federated learning frameworks typically follow the basic workflow of classical federated learning, with the Federated Averaging (FPAG) algorithm being the most representative aggregation method. Its workflow is as follows: the server initializes the global model and distributes it to the clients; each client trains using local data and uploads model updates (such as quantum gate rotation angle parameters); the server performs a weighted average of all updates to generate a new global model.

[0006] To alleviate communication pressure in federated learning, researchers have proposed various model compression techniques and attempted to apply them to QFL scenarios, mainly including model pruning, parameter quantization, and asynchronous updates. 1. Model Pruning: Reducing model size by removing unimportant parameters. Classic methods include L2 norm-based pruning, which considers parameters with smaller amplitudes to be less important. 2. Parameter Quantization: Converting high-precision floating-point parameters (e.g., 32-bit) into low-bit formats (e.g., 8-bit integers) for transmission. Classic methods include uniform quantization, which linearly maps the parameter range to a low-bit space.

[0007] While the aforementioned techniques alleviate the communication problem in classical federated learning to some extent, when they are directly applied to quantum federated learning, they reveal the following fundamental flaws due to their failure to consider the physical characteristics of quantum models:

[0008] (1) High-precision quantum parameters lead to huge communication overhead: Quantum model parameters (such as rotation angle and entanglement coefficient) are high-dimensional continuous variables and are extremely sensitive to quantum state evolution, requiring high-precision transmission. This results in a single round of communication data volume far exceeding that of classical neural networks, becoming the core bottleneck for the expansion of QFL systems.

[0009] (2) Classical compression methods severely damage the integrity of quantum structures: Static pruning based on classical metrics such as the L2 norm is physically unknowable and may incorrectly remove parameters with small amplitudes but crucial effects on quantum state evolution, thereby impairing the unitary nature of quantum circuits and weakening the expressive power of the model. Traditional low-bit quantization introduces significant quantization errors, equivalent to continuous parameter noise, which disrupts the precise quantum interference process and leads to a severe decline in model performance.

[0010] (3) Aggregation strategies fail to recognize the value of quantum information: Algorithms such as FedAvg only weight the data based on the amount of client data, completely ignoring the differences in the quality of quantum information contained in the updates of different client models. In heterogeneous data environments, some updates contain more quantum features that are useful to the global model, but existing aggregation strategies cannot identify and prioritize the fusion of these high-value information.

[0011] In conclusion, existing quantum federated learning communication optimization methods have failed to effectively alleviate communication load and require further optimization and improvement. Summary of the Invention

[0012] Therefore, the technical problem to be solved by the present invention is to overcome the problem that existing quantum federated learning communication optimization methods have failed to effectively alleviate communication load.

[0013] To address the aforementioned technical problems, this invention provides a communication optimization method for quantum federated learning, comprising:

[0014] Step S1: In quantum federated learning, the communication overhead generated by the transmission of data between the client and the server during the training process is modeled and the constraints are determined. The optimization objective is determined based on the modeled communication overhead and constraints.

[0015] Step S2: Based on the optimization objective, each client receives the global model from the server and then trains the global model for the current round.

[0016] Step S3: After each client completes the current round of training of the global model, it quantizes the format of the global model parameters obtained by training through a dynamic bit-width quantization strategy, and then uploads the quantized global model parameters to the server.

[0017] Step S4: The server receives the quantized global model parameters uploaded by each client, constructs the final aggregate weights corresponding to the global model parameters of each client based on the quantum Fisher information (QFI), aggregates the global model parameters of each client based on the final aggregate weights, and updates the global model based on the aggregated parameters to obtain the updated global model.

[0018] Step S5: The server dynamically prunes the updated global model based on the quantum Fisher information (QFI), removes redundant parameters, and quantizes the format of the global model parameters retained after pruning using a dynamic bit-width quantization strategy to generate a compressed model. The compressed global model is then distributed to each client.

[0019] Step S6: Repeat steps S2 to S5 until the preset training rounds are completed.

[0020] In one embodiment of the present invention, step S1 models the communication overhead generated during data transmission between the client and the server during the training process and determines the constraints. The method for determining the optimization objective based on the modeled communication overhead and constraints includes:

[0021] Model the communication overhead incurred during data transmission between the client and server during training:

[0022] In all rounds of communication during the training process, the cumulative communication volume is calculated based on the downlink communication overhead generated by the server distributing the global model to several clients, and the uplink communication overhead generated by several clients uploading the global model parameters trained in each round to the server. :

[0023] (1);

[0024] in, Total number of training rounds; Indicates client In the Uplink traffic of the wheel; For the client In the The number of parameters after round quantization; For the client In the Round-quantized model parameters; Indicates the server is in the Downlink traffic of the wheel; Indicates the first The number of parameters after round quantization; For the server in the Round-quantized model parameters; For the first The quantization bit width used by the wheel; The number of bits required to transmit metadata; For the number of clients;

[0025] To ensure the predictive power of the global model, the performance constraints of the global model are constructed as follows:

[0026] (2);

[0027] in, This represents the global model classification accuracy. These are the original global model parameters. These are the quantized model parameters. The allowable accuracy loss threshold;

[0028] Define the optimization goal:

[0029] (3);

[0030] in, Let cross-entropy loss function be defined as follows: ; For the client Local dataset; For the model to input The predicted output; Weighting factors are used to balance model performance and communication efficiency.

[0031] In one embodiment of the present invention, the dynamic bit-width quantization strategy includes: linear quantization and parameter normalization, bit packing mechanism, and dynamic bit-width adjustment, wherein,

[0032] The linear quantization and parameter normalization are specifically as follows:

[0033] Obtain the parameter set before each round of parameter transmission. ,in To preserve the first global model after server pruning A floating-point parameter or client The parameter array obtained during training, Calculate the parameter set for the number of clients or the number of parameter arrays. minimum value With the maximum value and define the scaling factor. The process of quantizing floating-point parameters into fixed-width unsigned integers in a linear manner is represented as follows:

[0034] (4);

[0035] (5);

[0036] in, Indicates the current quantization bit width; The quantized parameters are represented by integers. Here is the quantization function, representing the rounding operation;

[0037] The bit packing mechanism is specifically as follows:

[0038] Multiple linearly quantized and parameter-normalized values ​​are concatenated into a single byte to form a bitstream. The final quantization result is represented in the form of a communication message packet as follows:

[0039] (6);

[0040] in, The quantization results include bitstream and metadata; For bitstream; metadata = { , , , } is used to record key solution quantization parameters;

[0041] The dynamic bit width control specifically refers to:

[0042] Let the initial bit width be The minimum bit width is In the In round-robin communication, quantization bit width satisfy:

[0043] (7).

[0044] In one embodiment of the present invention, the method for constructing the final aggregated weights corresponding to the global model parameters of each client based on the Quantum Fisher Information (QFI) in step S3, aggregating the global model parameters of each client based on the final aggregated weights, and updating the global model based on the aggregated parameters to obtain the updated global model includes:

[0045] The server calculates the Quantum Fisher Information (QFI) score for each client. Afterwards, Normalization ,in, To The normalization result; The maximum value of the Quantum Fisher Information (QFI) score across all clients;

[0046] Then, combine the weights based on the amount of data from each client. , For the client Local dataset, To sum the data sets from all clients to obtain the total data set size, the final aggregate weight for each client is calculated. Based on the final aggregate weight The system aggregates the parameters of the global model for each client, updates the global model based on the aggregated parameters, and obtains the updated global model.

[0047] In one embodiment of the present invention, the method for the server to dynamically prune the updated global model based on the quantum Fisher information (QFI) in step S4 includes:

[0048] According to the Quantum Fisher Information QFI score Build parameters Composite Importance Score :

[0049] (10);

[0050] in, This is the adjustment coefficient; For all parameters Quantum Fisher Information QFI Score A set; This is the set of trainable parameters for the global model on the client side.

[0051] Dynamic pruning rate for building a global model:

[0052] (11);

[0053] in, The dynamic pruning rate of the global model; and These represent the initial and final pruning rates, respectively. This is the current training round; Total number of training rounds;

[0054] according to and Constructing a binary pruning mask Through binary pruning mask Implement parameters of the global model Pruning:

[0055] (12);

[0056] in, These are the parameters after pruning; This is element-wise multiplication; This is a quantile function that calculates the threshold for a given percentage.

[0057] In one embodiment of the present invention, the quantum Fisher information QFI score The formula is:

[0058] (9);

[0059] in, For parameters Quantum Fisher Information QFI Score For the quantum Fisher information (QFI) evaluation sample set, For measurement operators; The input sample is obtained by parameterized quantum circuit. The generated quantum state; This is the set of trainable parameters for the global model on the client side. Represents measurement operators The expected value, representing the parameter Physical sensitivity to the output quantum state.

[0060] In one embodiment of the invention, the step of basing the final aggregation weight... To aggregate the parameters of the global model for each client, the formula is as follows:

[0061] (13);

[0062] in, This refers to the number of clients.

[0063] To address the aforementioned technical problems, this invention provides a communication optimization system for quantum federated learning, comprising:

[0064] Building Module: Used in quantum federated learning to model the communication overhead generated by data transmission between the client and server during the training process, determine the constraints, and determine the optimization objective based on the modeled communication overhead and constraints;

[0065] Training module: Based on the optimization objective, each client receives the global model from the server and trains the global model for the current round.

[0066] Quantization module: After each client completes the current round of training on the global model, it quantizes the format of the global model parameters obtained through training using a dynamic bit-width quantization strategy, and then uploads the quantized global model parameters to the server.

[0067] Aggregation module: The server receives the quantized global model parameters uploaded by each client, constructs the final aggregation weights corresponding to the global model parameters of each client based on the Quantum Fisher Information (QFI), aggregates the global model parameters of each client based on the final aggregation weights, and updates the global model based on the aggregated parameters to obtain the updated global model.

[0068] The pruning module is used by the server to dynamically prune the updated global model based on the quantum Fisher information (QFI), remove redundant parameters, and quantize the format of the global model parameters retained after pruning through a dynamic bit-width quantization strategy to generate a compressed model, which is then distributed to each client.

[0069] Judgment module: Used to repeat the process from the training module to the pruning module until the preset training rounds are completed.

[0070] To address the aforementioned technical problems, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the communication optimization method for quantum federated learning described above.

[0071] To address the aforementioned technical problems, the present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the communication optimization method for quantum federated learning described above.

[0072] The technical solution of the present invention has the following advantages compared with the prior art:

[0073] The communication optimization method for quantum federated learning constructed in this invention introduces quantum Fisher information (QFI)-driven parameter importance assessment, progressive pruning and dynamic bit-width quantization mechanisms, and a quantum Fisher information (QFI)-aware weighted aggregation strategy. This invention brings significant improvements to quantum federated learning in terms of communication efficiency, model performance and system adaptability.

[0074] This invention effectively solves the key bottlenecks of existing quantum federated learning in terms of communication efficiency, model accuracy, and system adaptability, and provides solid technical support for realizing an efficient, reliable, and scalable quantum distributed learning system. Attached Figure Description

[0075] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0076] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0077] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0078] Example 1

[0079] Reference Figure 1 As shown, this invention relates to a communication optimization method for quantum federated learning, comprising:

[0080] Step S1: In quantum federated learning, the communication overhead generated by the transmission of data between the client and the server during the training process is modeled and the constraints are determined. The optimization objective is determined based on the modeled communication overhead and constraints.

[0081] Step S2: Based on the optimization objective, each client receives the global model (Global Quantum Convolutional Neural Network (QCNN) model) from the server and then trains the global model for the current round.

[0082] Step S3: After each client completes the current round of training of the global model, it quantizes the format of the global model parameters obtained by training through a dynamic bit-width quantization strategy, and then uploads the quantized global model parameters to the server.

[0083] Step S4: The server receives the quantized global model parameters uploaded by each client, constructs the final aggregate weights corresponding to the global model parameters of each client based on the quantum Fisher information (QFI), aggregates the global model parameters of each client based on the final aggregate weights, and updates the global model based on the aggregated parameters to obtain the updated global model.

[0084] Step S5: The server dynamically prunes the updated global model based on the quantum Fisher information (QFI), removes redundant parameters, and quantizes the format of the global model parameters retained after pruning using a dynamic bit-width quantization strategy to generate a compressed model. The compressed global model is then distributed to each client.

[0085] Step S6: Repeat steps S2 to S5 until the preset training rounds are completed.

[0086] It is worth mentioning that, in this embodiment, before executing steps S2 to S6, the communication overhead generated by the transmission of data between the client and the server during the training process is modeled, and constraints and optimization objectives are determined in order to reduce the communication load.

[0087] The following is a detailed description of this embodiment:

[0088] The overall idea of ​​this invention is to construct a closed-loop optimization system of "sensing-compression-regulation," using quantum Fisher information (QFI) as the core metric for the physical sensitivity of parameters, to drive the adaptive adjustment of pruning, quantization, and aggregation processes, thereby achieving a shift from "blind compression" to "physically sensed compression." The following will describe in detail the problem modeling, strategy implementation details, and functional advantages.

[0089] I. Problem Modeling

[0090] The core challenge of quantum federated learning lies in the fact that the high-precision transmission of high-dimensional continuous quantum parameters (such as rotation angle and entanglement coefficients) results in a massive amount of data per round of communication, severely limiting the scalability of the system in resource-constrained networks. To systematically address this issue, this embodiment first formally models the communication optimization problem in quantum federated learning, clarifying the optimization objective and constraints.

[0091] Suppose that the quantum federated learning system includes One client, each client Holding local private datasets They will collaboratively train a global quantum convolutional neural network (QCNN) model. Indicates the first The client in the first Parameters of the model after local training Indicates the server is in the The global model parameters held by the round.

[0092] (1) Communication overhead modeling:

[0093] In each round of communication, the client compresses the updated parameters and uploads them to the server. The server then aggregates and distributes the compressed global model. To quantify the overall communication burden, a cumulative communication volume is defined. as follows:

[0094] (1)

[0095] in, Total number of training rounds; Indicates client In the Uplink traffic of the wheel; For the client In the The number of parameters after round quantization; For the client In the Round-quantized model parameters; Indicates the server is in the Downlink traffic of the wheel; Indicates the first The number of parameters after round quantization; For the server in the Round-quantized model parameters; For the first The quantization bit width used by the wheel; For transmitting metadata For the first The quantization bit width used by the wheel; The number of bits required to transmit metadata (such as scaling factor, minimum value, etc.); This refers to the number of clients.

[0096] (2) Model performance constraints:

[0097] To ensure that the model's predictive ability does not degrade during the optimization process, the following accuracy constraints are applied:

[0098] (2)

[0099] in, This represents the global model classification accuracy. These are the original global model parameters (parameters corresponding to the global model without any optimization). These are the compressed model parameters. The threshold for allowable accuracy loss.

[0100] (3) Joint optimization objective:

[0101] The core of this invention is to construct an optimization objective by minimizing the weighted sum of the global loss function and communication overhead while satisfying accuracy constraints.

[0102] (3)

[0103] Formula (3) is subject to formulas (1) and (2). Let cross-entropy loss function be defined as follows: ; For the client Local dataset; For the model to input The predicted output; Weighting factors are used to balance model performance and communication efficiency.

[0104] By modeling the problem as described above, this invention transforms the communication optimization problem of quantum federated learning into a constrained multi-objective optimization problem, and provides a theoretical basis and quantitative evaluation foundation for subsequent methods.

[0105] II. Dynamic Bit Width Quantization Strategy

[0106] To effectively reduce the communication burden caused by the transmission of high-dimensional continuous parameters in quantum federated learning, this embodiment proposes a dynamic bit-width quantization strategy. This strategy integrates linear quantization with parameter normalization, bit packing mechanism, and dynamic bit-width adjustment, significantly reducing the amount of communication data while ensuring model accuracy.

[0107] (1) Linear quantization and parameter normalization

[0108] Obtain the parameter set before each round of parameter transmission. ,in To preserve the first global model after server pruning A floating-point parameter, or the client The parameter array obtained during training, Calculate the parameter set for the number of clients or the number of parameter arrays. minimum value With the maximum value and define the scaling factor. The process of quantizing floating-point parameters into fixed-width unsigned integers in a linear manner is represented as follows:

[0109] (4)

[0110] (5)

[0111] in, Indicates the current quantization bit width; The quantized parameters are represented by integers. `x` is the quantization function, representing the rounding operation. This process maps floating-point numbers to... The fixed-point interval ensures that the error after data compression is within a controllable range, thereby achieving effective discretization and unification of model parameters.

[0112] (2) Bit Packing Mechanism

[0113] Multiple linearly quantized and parameter-normalized values ​​are concatenated into a single byte using a bit-filling buffer to form a bit stream. The final quantization result is represented in the form of a communication message packet as follows:

[0114] (6)

[0115] in, The quantization results include bitstream and metadata; For bitstream; metadata = { , , , This is used to record key solution quantization parameters. Based on this information, the original floating-point parameters are recovered with high precision, ensuring the consistency of model reconstruction.

[0116] (3) Dynamic bit width control

[0117] During federated training, the model's requirements for accuracy and compression ratio vary at different stages. To balance convergence stability and communication efficiency, this invention designs a round-driven dynamic bit-width adjustment mechanism. Let the initial bit width be... The minimum bit width is In the In round-robin communication, dynamic bit width The calculation formula is:

[0118] (7)

[0119] In the early stages of training, a higher bit width (such as 8 bits) is used to ensure model convergence and numerical stability; as the number of training rounds increases, the bit width is gradually reduced, thereby significantly improving the compression ratio in the later stages.

[0120] (4) Dequantization and recovery

[0121] Before server-side aggregation or client-side continued training, the system performs dequantization to restore integer parameters to floating-point form. The restoration formula is as follows:

[0122] (8)

[0123] During the dequantization process, the server utilizes metadata The parameters recorded in the data ensure the reversibility and error control of the quantization-dequantization process.

[0124] III. Gradual Pruning Strategy

[0125] To further reduce communication load and improve the robustness of federated systems in heterogeneous environments, this embodiment proposes a progressive pruning strategy guided by quantum Fisher information (QFI). This strategy dynamically evaluates the importance of model parameters using physically interpretable quantum Fisher information (QFI), constructs a composite score by combining the L2 norm of the parameters, and then gradually removes redundant parameters using a dynamically increasing pruning rate, achieving a balance between structural compression and performance preservation.

[0126] (1) Parameter importance assessment based on quantum Fisher information (QFI)

[0127] Quantum Fisher information (QFI) is a core tool in quantum statistical theory for measuring the estimability of parameters. Its physical significance lies in quantifying the sensitivity of a quantum state to parameter perturbations. This invention employs its diagonal approximation to reduce computational complexity. Specifically, for any parameter... Its quantum Fisher information (QFI) can be automatically differentiated through a corresponding circuit to calculate its gradient squared with respect to the expected value of the measurement, and then averaged over multiple training samples to obtain:

[0128] (9)

[0129] in, For parameters Quantum Fisher Information QFI Score For the quantum Fisher information (QFI) evaluation sample set, For measurement operators (such as Pauli-Z); The input sample is obtained by parameterized quantum circuit. The generated quantum state; This is the set of trainable parameters for the global model on the client side. for The Middle One trainable parameter; Represents measurement operators The expected value, representing the parameter Physical sensitivity to the output quantum state (it should be noted that...) It is a measurement operator. It is the expected value of the operator, and this index characterizes the parameters. Physical sensitivity to the output quantum state.

[0130] (2) Construction of composite importance score

[0131] Considering that quantum Fisher information (QFI) is significantly affected by quantum noise and measurement fluctuations, this embodiment fuses the quantum Fisher information (QFI) value with the L2 norm parameter to construct a parameter... Composite Importance Score :

[0132] (10)

[0133] in, This is the adjustment coefficient; For all parameters Quantum Fisher Information QFI Score The set of scores. This scoring mechanism effectively improves the robustness of pruning decisions while taking into account both quantum physics sensitivity and model parameter amplitude stability.

[0134] (3) Dynamic pruning rate and mask generation

[0135] To avoid performance loss due to excessive pruning in the early stages of the model, this invention employs a linear progressive pruning scheduling strategy. The pruning rate increases linearly with the number of training epochs, defined as follows:

[0136] (11)

[0137] in, The dynamic pruning rate of the global model; and These represent the initial and final pruning rates, respectively. This is the current training round; This refers to the total number of training rounds.

[0138] according to and Constructing a binary pruning mask Through binary pruning mask To compress the global model:

[0139] (12)

[0140] in, These are the parameters after pruning; This is element-wise multiplication; This is a quantile function used to calculate a threshold for a given percentage. This mask is applied to the parameter matrix to achieve structural sparsity, thus retaining only the weights with high importance. The pruned parameter set will then be further compressed in the quantization module.

[0141] IV. Weighted Aggregation Method for Quantum Fisher Information (QFI) Sensing

[0142] In the server-side aggregation process, this embodiment innovatively combines the weight of the client's local data volume with the normalized QFI score to construct a dual-weighted aggregation mechanism.

[0143] The server calculates the Quantum Fisher Information (QFI) score for each client. Then, normalize it to , for The normalization result; The maximum QFI score for Quantum Fisher Information across all clients is then calculated. This is followed by a weighted analysis of the data volume of each client. Calculate the final aggregate weight for each client. Based on the final aggregate weight The system aggregates the parameters of the global model for each client, updates the global model based on the aggregated parameters, and obtains the updated global model.

[0144] Based on the final aggregate weight To aggregate the parameters of the global model for each client, the formula is as follows:

[0145] (13)

[0146] This aggregation strategy prioritizes the fusion of client updates that have large data volumes and high quantum information quality (high QFI scores), thereby improving the convergence speed and generalization performance of the global model.

[0147] V. Dynamic Update Mechanism of Quantum Fisher Information (QFI) Sample Set

[0148] To ensure the timeliness and representativeness of the parameter sensitivity assessment, the server dynamically updates the Quantum Fisher Information (QFI) assessment sample set after each round of aggregation. To ensure the timeliness of the evaluation, this invention employs a stratified sampling strategy for the sample set. Dynamic updates are performed to adapt to changes in data distribution during model training. This mechanism ensures that the Quantum Fisher Information (QFI) score is always calculated based on the most representative subset of the current data, maintaining the adaptability and robustness of the entire framework.

[0149] The present invention will be described in detail below through a specific embodiment:

[0150] I. Construction and Training of a Quantum Federated Learning Communication Optimization System

[0151] This embodiment demonstrates the complete construction and training process of the QFI-Opt system in a simulated quantum federated learning environment.

[0152] 1. System environment configuration:

[0153] Hardware platform: Two Intel(R) Xeon(R) CPU E5-2685 v3 @ 2.60GHz processors (each processor has 12 cores) and two NVIDIA GeForce GTX 1080 Ti graphics cards, simulating multiple distributed quantum clients.

[0154] Software environment: The operating system is Linux, the quantum computing simulation framework is PennyLane, the federated learning simulation platform is Flower, and the deep learning framework is PyTorch.

[0155] Quantum backend: The PennyLane simulator is used, which supports 8-qubit quantum circuit simulation. Specific experimental hyperparameter settings are shown in Table 1.

[0156] Table 1 Experimental hyperparameters

[0157]

[0158] 2. Dataset and Model Preparation:

[0159] Datasets: The CIFAR-10 and Fashion-MNIST datasets were selected to construct a binary classification task (class 0 and class 1). The client data distribution was set to non-independent and identically distributed (Non-IID) using a Dirichlet distribution to simulate real-world heterogeneous scenarios.

[0160] Quantum Models: This embodiment employs three quantum convolutional neural network (QCNN) architectures as native models: (a) Standard QCNN (3 convolutional layers + 3 pooling layers, 51 parameters); (b) Pooling-Free QCNN (3 convolutional layers, 45 parameters); and (c) 1D Chain QCNN (linear structure, 45 parameters). All models encode classical data into quantum states using angular embedding, and the output is obtained through Pauli-Z expectation measurement.

[0161] 3. QFI-Opt model training process:

[0162] Step 1: Initialization. The server initializes the global QCNN model parameters using the Xavier strategy (a commonly used method for initializing neural network parameters) and constructs an initial quantum Fisher information (QFI) evaluation sample set (the number of samples is 20 by default) through hierarchical sampling.

[0163] Step 2: Local training on the client side. The server distributes the current global model to 10 clients. After receiving the model, if it is in quantized format, the client first performs dequantization (Equation (8)) to restore full precision. Subsequently, each client trains the global model. Each client uses local data to perform 10 local training epochs. The batch size is set to 64, and the optimizer uses stochastic gradient descent (SGD) with momentum and a learning rate of 0.01.

[0164] Step 3: Client-side compression and upload. After the client completes local training of the global model, it executes a dynamic bit-width quantization strategy to update the model parameters (formulas (4)-(6)): initial bit width =8 bits, minimum bit width =2 bits, the quantization process includes linear scaling and bit packing. Quantized model parameters and metadata { , , , Upload to the server.

[0165] Step 4: Server-side Quantum Fisher Information (QFI) Sensing Aggregation and Compression. The server first receives the quantization model parameters uploaded by each client and performs dequantization. Then, it uses the Quantum Fisher Information (QFI) sample set to calculate the average Quantum Fisher Information (QFI) score for each client. (Formula (9)). Next, the server combines the data volume weights. Compared with the normalized quantum Fisher information (QFI) score The parameter aggregation operation for information quality perception is performed (Formula (13)). The global model is then updated based on the aggregated parameters to generate an updated global model. Based on this, the server constructs a composite importance score using quantum Fisher information (QFI) and the L2 norm. (Formula (10)), where The value is set to 0.7. A pruning mask is generated based on the progressive pruning rate (Equation (11)), and redundant parameters are removed by pruning (Equation (12)). Finally, a dynamic bit-width quantization strategy (Equations (4)-(6)) is applied to the parameters retained after pruning to generate a compressed global model.

[0166] Step 5: Model Distribution and Sample Set Update. The server distributes the compressed global model to each client for retraining and dynamically updates the Quantum Fisher Information (QFI) evaluation sample set based on the current model state to maintain its representativeness.

[0167] Step 6: Iteration Loop. Repeat steps 2 through 5 until the preset number of global communication rounds is completed. =5 rounds).

[0168] 4. Performance Evaluation:

[0169] During training, the communication overhead (uplink / downlink bits), model accuracy and cross-entropy loss on the test set, client training time and its standard deviation are recorded in real time for each round.

[0170] Experimental results show that, under the standard QCNN architecture of the Fashion-MNIST dataset, the average communication overhead of this invention is reduced by 73.04% compared to the FedAvg baseline (specific communication data are shown in Tables 3a, 3b, and 3c, corresponding to the global models using Standard QCNN, Pooling-Free QCNN, and 1D Chain QCNN, respectively, comparing the communication overhead (number of bits) of different model structures and algorithms on the Fashion-MNIST dataset). Simultaneously, it maintains or surpasses the model accuracy of the compressed baseline (specific accuracy data are shown in Tables 2a, 2b, and 2c, corresponding to the global models using Standard QCNN, Pooling-Free QCNN, and 1D Chain QCNN, respectively, comparing the accuracy of different model structures and algorithms on the Fashion-MNIST dataset), verifying its efficiency and practicality.

[0171] Table 2a Standard QCNN Accuracy

[0172]

[0173] Table 2b Accuracy of Pooling-Free QCNN

[0174]

[0175] Table 2c 1D Chain QCNN Accuracy

[0176]

[0177] Table 3a Standard QCNN Communication Rate

[0178]

[0179] Table 3b Pooling-Free QCNN Communication Rate

[0180]

[0181] Table 3c 1D Chain QCNN Communication Volume

[0182]

[0183] A composite importance scoring mechanism, constructed by fusing quantum Fisher information (QFI) and the L2 norm, effectively mitigates fluctuations caused by quantum noise and client data heterogeneity (Non-IID), enhancing the stability of pruning and quantization processes. Furthermore, the QFI-aware aggregation strategy comprehensively considers both the amount of client data and the quality of quantum information, giving higher weight to high-value updates in the global model, thereby accelerating model convergence and improving generalization ability. Experiments show that QFI-Opt's training time standard deviation in heterogeneous client environments is significantly lower than that of traditional methods (for specific comparisons, see Tables 4a, 4b, and 4c for the average training time (s) of different model structures and algorithms (Standard QCNN, Pooling-Free QCNN, 1D Chain QCNN) on the Fashion-MNIST dataset, and Tables 5a, 5b, and 5c for the average training time (s) of different model structures and algorithms (Standard QCNN, Pooling-Free QCNN, 1D Chain QCNN) on the CIFAR-10 dataset), demonstrating excellent load balancing and system robustness.

[0184] Table 4a shows the average training time of Standard QCNN on the Fashion-MNIST dataset.

[0185]

[0186] Table 4b shows the average training time of Pooling-Free QCNN on the Fashion-MNIST dataset.

[0187]

[0188] Table 4c Average training time of 1D Chain QCNN on the Fashion-MNIST dataset

[0189]

[0190] Table 5a shows the average training time of Standard QCNN on the CIFAR-10 dataset.

[0191]

[0192] Table 5b shows the average training time of Pooling-Free QCNN on the CIFAR-10 dataset.

[0193]

[0194] Table 5c Average training time of 1D Chain QCNN on the CIFAR-10 dataset

[0195]

[0196] Example 2

[0197] This embodiment provides a communication optimization system for quantum federated learning, including:

[0198] Building Module: Used in quantum federated learning to model the communication overhead generated by data transmission between the client and server during the training process, determine the constraints, and determine the optimization objective based on the modeled communication overhead and constraints;

[0199] Training module: Based on the optimization objective, each client receives the global model from the server and trains the global model for the current round.

[0200] Quantization module: After each client completes the current round of training on the global model, it quantizes the format of the global model parameters obtained through training using a dynamic bit-width quantization strategy, and then uploads the quantized global model parameters to the server.

[0201] Aggregation module: The server receives the quantized global model parameters uploaded by each client, constructs the final aggregation weights corresponding to the global model parameters of each client based on the Quantum Fisher Information (QFI), aggregates the global model parameters of each client based on the final aggregation weights, and updates the global model based on the aggregated parameters to obtain the updated global model.

[0202] The pruning module is used by the server to dynamically prune the updated global model based on the quantum Fisher information (QFI), remove redundant parameters, and quantize the format of the global model parameters retained after pruning through a dynamic bit-width quantization strategy to generate a compressed model, which is then distributed to each client.

[0203] Judgment module: Used to repeat the process from the training module to the pruning module until the preset training rounds are completed.

[0204] Example 3

[0205] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the communication optimization method for quantum federated learning described in Embodiment 1.

[0206] Example 4

[0207] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the communication optimization method for quantum federated learning described in Embodiment 1.

[0208] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this 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 storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0209] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0210] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0211] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0212] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0213] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A communication optimization method for quantum federated learning, characterized in that: include: Step S1: In quantum federated learning, the communication overhead generated by the transmission of data between the client and the server during the training process is modeled and the constraints are determined. The optimization objective is determined based on the modeled communication overhead and constraints. Step S2: Based on the optimization objective, each client receives the global model from the server and then trains the global model for the current round. Step S3: After each client completes the current round of training of the global model, it quantizes the format of the global model parameters obtained by training through a dynamic bit-width quantization strategy, and then uploads the quantized global model parameters to the server. Step S4: The server receives the quantized global model parameters uploaded by each client, constructs the final aggregate weights corresponding to the global model parameters of each client based on the quantum Fisher information (QFI), aggregates the global model parameters of each client based on the final aggregate weights, and updates the global model based on the aggregated parameters to obtain the updated global model. Step S5: The server dynamically prunes the updated global model based on the quantum Fisher information (QFI), removes redundant parameters, and quantizes the format of the global model parameters retained after pruning using a dynamic bit-width quantization strategy to generate a compressed model. The compressed global model is then distributed to each client. Step S6: Repeat steps S2 to S5 until the preset training rounds are completed.

2. The communication optimization method for quantum federated learning according to claim 1, characterized in that: Step S1 involves modeling the communication overhead generated during data transmission between the client and server during training and determining the constraints. The method for determining the optimization objective based on the modeled communication overhead and constraints includes: Model the communication overhead incurred during data transmission between the client and server during training: In all rounds of communication during the training process, the cumulative communication volume is calculated based on the downlink communication overhead generated by the server distributing the global model to several clients, and the uplink communication overhead generated by several clients uploading the global model parameters trained in each round to the server. : （1）； in, Total number of training rounds; Indicates the client In the Uplink traffic of the wheel; For the client In the The number of parameters after round quantization; For the client In the Round-quantized model parameters; Indicates the server is in the Downlink traffic of the wheel; Indicates the first The number of parameters after round quantization; For the server in the Round-quantized model parameters; For the first The quantization bit width used by the wheel; The number of bits required to transmit metadata; For the number of clients; To ensure the predictive power of the global model, the performance constraints of the global model are constructed as follows: （2）； in, This represents the global model classification accuracy. These are the original global model parameters. These are the quantized model parameters. The allowable accuracy loss threshold; Define the optimization goal: （3）； in, Let cross-entropy loss function be defined as follows: ; For the client Local dataset; For the model to input The predicted output; Weighting factors are used to balance model performance and communication efficiency.

3. The communication optimization method for quantum federated learning according to claim 1, characterized in that: The dynamic bit-width quantization strategy includes: linear quantization and parameter normalization, bit packing mechanism, and dynamic bit-width adjustment, wherein... The linear quantization and parameter normalization are specifically as follows: Obtain the parameter set before each round of parameter transmission. ,in To preserve the first global model after server pruning A floating-point parameter or client The parameter array obtained during training, Calculate the parameter set for the number of clients or the number of parameter arrays. minimum value With the maximum value and define the scaling factor. The process of linearly quantizing floating-point parameters into fixed-width unsigned integers is represented as follows: (4)； (5)； in, Indicates the current quantization bit width; The quantized parameters are represented by integers. Here is the quantization function, representing the rounding operation; The bit packing mechanism is specifically as follows: Multiple linearly quantized and parameter-normalized values ​​are concatenated into a single byte to form a bitstream. The final quantization result is represented in the form of a communication message packet as follows: (6)； in, The quantization results include bitstream and metadata; For bitstream; metadata = { , , , } is used to record dequantization parameters; The dynamic bit width control specifically refers to: Let the initial bit width be The minimum bit width is In the In round-robin communication, quantization bit width satisfy: (7)。 4. The communication optimization method for quantum federated learning according to claim 1, characterized in that: In step S3, the final aggregated weights corresponding to the global model parameters of each client are constructed based on the Quantum Fisher Information (QFI). The global model parameters of each client are aggregated based on these final aggregated weights. The global model is then updated based on the aggregated parameters to obtain the updated global model. The method includes: The server calculates the Quantum Fisher Information (QFI) score for each client. Afterwards, Normalization ,in, To The normalization result; The maximum value of the Quantum Fisher Information (QFI) score across all clients; Then, combine the weights based on the amount of data from each client. , For the client Local dataset, To sum the data sets from all clients to obtain the total data set size, the final aggregate weight for each client is calculated. Based on the final aggregate weight The system aggregates the parameters of the global model for each client, updates the global model based on the aggregated parameters, and obtains the updated global model.

5. The communication optimization method for quantum federated learning according to claim 4, characterized in that: The method for the server to dynamically prune the updated global model based on the quantum Fisher information (QFI) in step S4 includes: According to the Quantum Fisher Information QFI score Build parameters Composite Importance Score : (10)； in, This is the adjustment coefficient; For all parameters Quantum Fisher Information QFI Score A set; This is the set of trainable parameters for the global model on the client side. Dynamic pruning rate for building a global model: (11)； in, The dynamic pruning rate of the global model; and These represent the initial and final pruning rates, respectively. This is the current training round; Total number of training rounds; according to and Constructing a binary pruning mask Through binary pruning mask Implement parameters of the global model Pruning: (12)； in, These are the parameters after pruning; This is element-wise multiplication; This is a quantile function that calculates the threshold for a given percentage.

6. The communication optimization method for quantum federated learning according to claim 4 or 5, characterized in that: The quantum Fisher information QFI score The formula is: (9)； in, For parameters Quantum Fisher Information QFI Score For the quantum Fisher information (QFI) evaluation sample set, For measurement operators; The input sample is obtained by parameterized quantum circuit. The generated quantum state; This is the set of trainable parameters for the global model on the client side. Represents measurement operators The expected value, representing the parameter Physical sensitivity to the output quantum state.

7. The communication optimization method for quantum federated learning according to claim 5, characterized in that: The final aggregation weight is used as the basis for the calculation. To aggregate the parameters of the global model for each client, the formula is as follows: (13)； in, This refers to the number of clients.

8. A communication optimization system for quantum federated learning, characterized in that: include: Building Module: Used in quantum federated learning to model the communication overhead generated by data transmission between the client and server during the training process, determine the constraints, and determine the optimization objective based on the modeled communication overhead and constraints; Training module: Based on the optimization objective, each client receives the global model from the server and trains the global model for the current round. Quantization module: After each client completes the current round of training on the global model, it quantizes the format of the global model parameters obtained through training using a dynamic bit-width quantization strategy, and then uploads the quantized global model parameters to the server. Aggregation module: The server receives the quantized global model parameters uploaded by each client, constructs the final aggregation weights corresponding to the global model parameters of each client based on the Quantum Fisher Information (QFI), aggregates the global model parameters of each client based on the final aggregation weights, and updates the global model based on the aggregated parameters to obtain the updated global model. The pruning module is used by the server to dynamically prune the updated global model based on the quantum Fisher information (QFI), remove redundant parameters, and quantize the format of the global model parameters retained after pruning through a dynamic bit-width quantization strategy to generate a compressed model, which is then distributed to each client. Judgment module: Used to repeat the process from the training module to the pruning module until the preset training rounds are completed.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the communication optimization method for quantum federated learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the communication optimization method for quantum federated learning as described in any one of claims 1 to 7.

Images