A perception quantization incentive method, device and system for communication efficient federated learning

Abstract

The application discloses a kind of perception quantization incentive methods, devices and systems for communication efficient federal learning.The method of the present application allows clients to submit bids based on different quantization strategies, by introducing predetermined parameters to approximate the originally difficult to handle constraints, the interdependent goals of global social welfare maximization and payment rule design are decoupled into subproblems that can be sequentially solved;In the node selection stage, the constraints are incorporated into the objective function using the Lagrange dual method;In the payment stage, based on the VCG principle, design the preliminary payment rule, and use quadratic programming to scale to meet the actual budget constraints.The present application realizes approximate dominant strategy incentive compatibility (-DSIC) within polynomial time complexity, which can effectively improve the convergence accuracy of the global model within the limited budget, and significantly reduce the communication time.

Description

Technical Field

[0001] This invention belongs to the field of digital information transmission technology, and relates to federated learning and incentive mechanism technology, specifically to a perceptual quantization incentive method, device and system for efficient federated learning in communication. Background Technology

[0002] Federated Learning (FL) has emerged as a promising privacy-preserving distributed machine learning paradigm. Within this framework, multiple distributed clients train models locally using their own data and only upload model updates (gradients) to a central server, rather than the original data. This significantly reduces the privacy risks common in traditional centralized machine learning methods.

[0003] However, participating in federated learning incurs significant computational and communication costs for clients. Because model training and data transfer are resource-intensive, clients with high computing power and high-quality data are often reluctant to actively participate in collaborative training to avoid "free-riding" by other participants. Therefore, designing an efficient and effective incentive mechanism to encourage high-quality clients to join the federated learning system is crucial.

[0004] In federated learning, as the number of parameters in deep learning models increases, the communication overhead between clients and servers rises sharply. To address this communication bottleneck, quantization techniques have become one of the mainstream solutions. Quantization not only effectively compresses the gradient size uploaded by each client and reduces communication overhead, but also alleviates the "lagging" problem caused by bandwidth limitations in heterogeneous communication networks.

[0005] Although quantization techniques are widely used in federated learning, existing incentive mechanisms (primarily based on game theory, contract theory, or auction mechanisms) mainly focus on evaluating the client's "training quality," such as the size of the training data, the quality of the data distribution, and the training duration. These existing mechanisms generally ignore the impact of communication compression (such as quantization) during model updates. This leads to the server's inability to accurately assess and characterize the true quality of the quantization gradients it receives in federated learning scenarios that employ compression techniques like quantization. When clients dynamically adjust their quantization levels or adopt different quantization strategies, the lack of effective assessment of the quantization impact prevents existing mechanisms from accurately measuring each client's true contribution, ultimately resulting in system incentive misalignment and severe unfair resource allocation, thus weakening the effectiveness of the incentive mechanism.

[0006] Therefore, there is an urgent need for a new incentive mechanism that can sense the level of quantification, support multi-strategy bidding, and maximize social welfare (i.e., maximize the sum of benefits or utility of all participants) under a limited budget. Summary of the Invention

[0007] Purpose of the invention: To address the problems existing in the prior art, this invention proposes a Quantization-Aware Incentive Mechanism (QUAIM), device, and system for efficient federated learning. Based on a reverse auction architecture, it enables federated learning systems to effectively evaluate the impact of quantization strategies on the quality of client updates and optimize client selection and payment allocation under conditions of limited communication and budget.

[0008] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0009] A perceptual quantization incentive method for efficient federated learning is implemented on the server side of the federated learning process, and includes the following steps:

[0010] (1) Receive various quantitative strategies and their corresponding bidding information submitted by each client;

[0011] (2) Based on the fitted quantization influence function, evaluate the model update quality of each client at different quantization levels;

[0012] (3) Replace the total budget constraint with the total bid limit of the selected clients to achieve problem decoupling. In the node selection stage, use the Lagrange dual method to construct the objective function and solve for the selected optimal clients and the set of quantization strategies.

[0013] (4) Determine the initial payment rules that satisfy the strict dominance strategy incentive compatibility according to the VCG mechanism, and then scale and adjust the initial payment amount through quadratic programming to meet the budget constraint conditions, and calculate the final amount to be paid to each selected client.

[0014] Specifically, in step (2), the formula for the quantification influence function is as follows:

[0015] ,

[0016] in, This indicates the specific quantization bit width under this strategy, parameter These are the parameters adjusted through fitting;

[0017] The updated quality is calculated using the following formula:

[0018] ,

[0019] in, Let i be the amount of training data for client i. This is the loss value under full precision without quantization.

[0020] Specifically, in step (3), the objective function constructed using the Lagrange duality method is as follows:

[0021]

[0022] in, For Lagrange multipliers, To update quality, The client bids the price. This is the budget constraint for the current round.

[0023] Specifically, in step (4), the preliminary payment rule calculation formula is as follows:

[0024] ,

[0025] in, This indicates that the server is on the client side. Total update quality collected during participation Indicates virtual culling client The total quality of updates that the system can collect afterward To update quality, The bid price for the client.

[0026] Specifically, in step (4), the formula for scaling the initial payment amount using quadratic programming is as follows:

[0027] ,

[0028] in, For preliminary payment rules, This is the final payment rule.

[0029] Specifically, in step (4), the budget constraints include:

[0030] ,

[0031] ,

[0032] in, Due to budget constraints for the current round, For the final payment rules, For the selected client and its set of policies, Costs for the client.

[0033] Furthermore, the following steps are included before each of the previous steps:

[0034] At the start of each round of federated learning, the server broadcasts the current global model parameters to all participating candidate clients; the server can receive parameters that the clients obtain after receiving the model and training it locally, which can characterize the local data distribution and update potential.

[0035] Furthermore, the following steps are included after each step:

[0036] The server sends a selection notification and the assigned quantization strategy to the selected client;

[0037] Subsequently, the following process is performed on the client: the selected client performs stochastic gradient descent model training on its local machine to obtain updated gradients, and performs quantization compression operation on the local gradient according to the quantization strategy bit width obtained in step (3), and finally uploads the compressed and quantized gradient matrix to the server.

[0038] The present invention also provides a perceptual quantization excitation device for efficient federated learning in communication, comprising:

[0039] The bidding receiving module receives various quantitative strategies and their corresponding bidding information submitted by each client.

[0040] The quality assessment module is used to evaluate the quality of model updates for each client at different quantization levels using a quantization impact function.

[0041] The Lagrange node selection module is used to replace the total budget constraint with the sum of bids from selected clients, thereby decoupling the problem. In the node selection phase, the objective function is constructed using the Lagrange duality method, and the optimal selected client and quantization strategy set are solved.

[0042] The secondary planning payment scaling module determines the initial payment rules that satisfy the strict dominance strategy incentive compatibility based on the VCG mechanism, and then scales and adjusts the initial payment amount through secondary planning to meet the budget constraints, and calculates the final amount to be paid to each selected client.

[0043] The present invention also provides a computer system, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the perceptual quantization incentive method for efficient federated learning of communication as described above.

[0044] Beneficial effects:

[0045] This invention not only achieves accurate evaluation and dynamic resource allocation of client contributions in quantized federated learning, ensuring deployment feasibility in large-scale real-world systems, but also satisfies approximate dominance policy incentive compatibility within polynomial time complexity through a quadratic programming scaling mechanism. (-DSIC) ensures that the additional benefits that clients can gain through false reporting costs are strictly limited, and that these benefits asymptotically approach zero as the number of participants increases. Experimental results show that this invention can collect higher-quality model updates within a given budget. Compared with existing benchmark methods, it not only significantly improves the global model convergence accuracy but also greatly reduces communication time, while ensuring compatibility between individual rationality and approximate incentives. Attached Figure Description

[0046] Figure 1 This is a framework for an incentive mechanism based on reverse auctions.

[0047] Figure 2 A schematic diagram of the perceptual quantization incentive method for efficient federated learning in communication provided by the present invention;

[0048] Figure 3 The graph shows the quantized influence function obtained from the fitting process.

[0049] Figure 4 This is a comparison chart of the accuracy of the method of the present invention (QUAIM) with other common federated learning methods in the field;

[0050] Figure 5 This is a comparison chart of communication time between the method of the present invention (QUAIM) and other common federated learning methods in the field;

[0051] Figure 6 A scatter plot illustrating the additional utility gained by forcing some clients to lie about their quotes using the method of this invention. Detailed Implementation

[0052] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0053] A typical federated learning system consists of a central server and multiple distributed clients. Each client has a local dataset and exhibits heterogeneity in local computing resources and network bandwidth. The entire mechanism operates within a pre-set fixed budget. Within this framework, the training process of federated learning is optimized by maximizing the total social welfare of the system (i.e., the total model update quality collected by the server minus the total payment cost). Specifically, the perceptual quantization incentive method for efficient communication-oriented federated learning provided by this invention is executed on the server side of the federated learning, and the implementation framework is as follows: Figure 1 As shown, the process in a complete global iteration of federated learning is as follows: Figure 2 As shown, the specific steps include the following:

[0054] Step S1, Global Model Distribution and Update Quality Prediction: In each round (the... At the start of a round of federated learning, the server broadcasts the current global model parameters to all participating candidate clients. After receiving the model, each client performs a small amount of pre-training using its local data and uploads parameters that characterize its local data distribution and update potential (such as loss variation and data size) to the server. Figure 2 As shown on the left. The server has a built-in Quantization Impact Function, the graph of which is shown below. Figure 3 As shown, quantization behavior can be evaluated without the client actually transmitting a huge, complete gradient. In this embodiment, the server uses a hyperbolic tangent function to fit the effect of the quantization bit width on the model loss. For the client... The adopted first The impact coefficient of various quantitative strategies The calculation formula is:

[0055] ,

[0056] in, This indicates the specific quantization bit width under this strategy, parameter This can be obtained through fitting the data in a few early cold start rounds. Using this function, after receiving the quantization strategy uploaded by the client in step S2, the server can efficiently and accurately predict the final update quality of the client under different quantization strategies. The specific calculation formula is as follows:

[0057] ,

[0058] in, Let i be the amount of training data for client i. This is the loss value under full precision without quantization.

[0059] Step S2, Client-side multi-strategy bidding reception: Due to limited communication resources, clients can choose different quantization bit widths (such as 4-bit, 8-bit, 16-bit, etc.) to compress the transmission gradient. Each client... It comprehensively assesses its own computational and network transmission costs, and targets each quantization strategy it supports. Propose the corresponding bid price The server receives a set of data uploaded by various clients, containing multiple quantitative strategies and their corresponding bidding prices, and then proceeds to the reverse auction phase, such as... Figure 2 As shown on the right. This approach allows clients to submit diverse bidding information to the server based on different quantification strategies (i.e., different communication costs for different quantification levels). This enables finer-grained resource allocation while reducing overall communication overhead.

[0060] Step S3, Node and Strategy Selection Based on Lagrange Decoupling: The optimization objective of the server is to achieve the following within the budget. The optimal client and its quantification strategy combination are selected internally. In the design of the reverse auction mechanism, since the actual "budget constraint" depends on the "final payment rule," and the "final payment rule" itself is calculated based on the "node selection result," there is a strong interdependent coupling relationship between the two, which is an NP-hard problem that is difficult to solve precisely in polynomial time. To decouple this problem, this embodiment of the invention introduces the client's bidding price. As a known intermediate variable, the original constraint "the total payment amount of the server shall not exceed the budget" is approximately replaced with "the total bidding amount of the selected client under its selected quantitative strategy shall not exceed the budget," which requires... This innovation successfully decouples the interdependent complex problem into two sub-problems, "node selection" and "payment rule design," which can be solved sequentially. Based on this, in the node selection phase, the server employs the Lagrange duality method to construct the objective function:

[0061] ,

[0062] in, For Lagrange multipliers, Due to budget constraints in the current round, the server breaks down this massive global optimization problem into... Solve each independent client-side subproblem separately. By differentiating each subproblem and applying a binary search algorithm, the continuous quantization bit width that achieves a local optimum can be efficiently calculated. Subsequently, discretization is performed by rounding to the nearest integer to determine the final selected client indicator variable (if selected). (or 0 if the condition is true, otherwise it is false) and the optimal quantization strategy for each selected client, forming the final selection set. The process is as follows Figure 2 The node selection section on the right is shown.

[0063] Step S4, Preliminary payment design based on VCG mechanism: Determining the selected client set. Afterwards, the server needs to formulate preliminary payment rules. This ensures the mechanism satisfies strict Dominant Strategy Incentive Compatibility (DSIC). This property is central to the auction mechanism design, mathematically guaranteeing that any client's "honest reporting of their true cost" is their optimal dominance strategy. The initial payout amount is calculated using the following formula:

[0064] ,

[0065] in, This indicates that the server is on the client side. Total update quality collected during participation Indicates virtual culling client The total quality of updates that the system can collect afterward. This formula ensures that the reward for the selected client is directly linked to its "marginal contribution" to the total quality of the system, as follows: Figure 2 The payment rules are shown on the right.

[0066] Step S5, Actual Payment Scaling Based on Quadratic Programming: Although the initial payment rules in Step S4... While a strict DSIC mechanism is guaranteed, the actual total payment amount after aggregation will inevitably exceed or deviate from the system's true budget. Therefore, the server applies a quadratic programming algorithm to reasonably scale and adjust the initial payment amount, ensuring it strictly meets the system's actual budget constraints and individual rationality constraints. The objective function of this quadratic programming problem is defined as minimizing the deviation between the actual payment and the initial DSIC payment:

[0067] ,

[0068] in, This indicates the initial payment rules under the VCG mechanism. This is the final payment rule.

[0069] At the same time, the following strict constraints must be met: (1) System budget constraints: That is, the total amount of rewards actually distributed after scaling cannot exceed the budget for this round of federated learning, of which, For the selected client and its set of policies. (2) Individual rationality constraint: This means that the final payment to each selected client must be greater than the actual cost of performing local training and quantization communication, thereby ensuring that the net utility of the participants is positive and that they have sufficient motivation to participate. Let's consider the client's cost. Solving this quadratic programming problem yields the actual payment amount for each selected client. The mechanism after this second scaling step successfully satisfies the incentive compatibility of the approximate dominant strategy. -DSIC). This means that even if a client attempts to profit through false quotes, there is a strict cap on the additional gains it can obtain, and this cap applies to the scale of clients participating in federated learning. At that time, the upper limit of this additional income .

[0070] Step S6: Local Training Execution, Quantization Communication, and Global Aggregation: The server sends a selection notification and the assigned quantization strategy to the selected clients. The selected clients execute the corresponding stochastic gradient descent (SGD) model training on their local devices to obtain updated gradients. They strictly adhere to the quantization strategy bit width assigned in Step S3, performing quantization compression on the local gradients, and finally uploading the compressed and quantized gradient matrix to the server. For unselected clients, their payment for this round is counted as 0, and no invalid redundant uploads are performed. After receiving all compressed gradients, the server decodes and restores them, and uses an aggregation algorithm (such as average aggregation) to aggregate the restored gradients and update the global model parameters. Finally, the server distributes the actual payment to the corresponding nodes according to the results calculated in Step S5, completing this round of federated learning iteration. This entire process is repeated continuously until the global model reaches the target convergence accuracy within a finite budget.

[0071] Finally, experiments verified the effectiveness of the perceptual quantification incentive method. Figure 4 This demonstrates that the method of the present invention achieves higher accuracy within a limited budget compared to other incentive mechanisms. Figure 5 This demonstrates that the method of the present invention can reduce communication time under different budget conditions. To prove the approximate incentive compatibility property of the method of the present invention, we forced 20% of the clients to act as "speculators" and make false bids, and the results are as follows. Figure 6 As shown, we can see that only 7.5% of clients gain additional utility from false reporting, and the additional utility gained is very low.

[0072] Based on the same ideas as the main steps in a perceptual quantization incentive method for efficient federated learning, the present invention also provides a perceptual quantization incentive device for efficient federated learning, comprising:

[0073] The bidding receiving module is used to receive bidding requests submitted by clients based on multiple quantitative strategies, which include quantitative strategies and bidding information.

[0074] The quality assessment module is used to evaluate the quality of model updates for each client at different quantization levels using a quantization impact function.

[0075] The Lagrange node selection module is used to replace the total budget constraint with the sum of bids from selected clients, thereby decoupling the problem and using the Lagrange dual method to find the optimal client and quantization strategy set.

[0076] The secondary planning payment scaling module calculates the initial payment amount based on the VCG mechanism, and then scales the amount through secondary planning to meet the system's total budget constraints and individual rationality conditions.

[0077] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a perceptual quantization incentive method for efficient federated learning in communication.

[0078] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a perceptual quantization incentive method for communication-efficient federated learning.

[0079] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.

[0080] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] 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.

[0082] 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.

[0083] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A perceptual quantization incentive method for efficient federated learning, executed on the server side of the federated learning, characterized in that, Includes the following steps: (1) Receive various quantitative strategies and their corresponding bidding information submitted by each client; (2) Based on the fitted quantization influence function, evaluate the model update quality of each client at different quantization levels; (3) Replace the total budget constraint with the total bid limit of the selected clients to achieve problem decoupling. In the node selection stage, use the Lagrange dual method to construct the objective function and solve for the selected optimal clients and the set of quantization strategies. (4) Determine the initial payment rules that satisfy the strict dominance strategy incentive compatibility according to the VCG mechanism, and then scale and adjust the initial payment amount through quadratic programming to meet the budget constraint conditions, and calculate the final amount to be paid to each selected client.

2. The perceptual quantization incentive method for efficient federated learning in communication according to claim 1, characterized in that, In step (2), the formula for the quantification influence function is as follows: ， in, This indicates the specific quantization bit width under this strategy, parameter These are the parameters adjusted through fitting; The updated quality is calculated using the following formula: ， in, Let i be the amount of training data for client i. This is the loss value under full precision without quantization.

3. The perceptual quantization incentive method for efficient federated learning in communication according to claim 1, characterized in that, In step (3), the objective function constructed using the Lagrange duality method is as follows: ， in, For Lagrange multipliers, To update quality, The client bids the price. This is the budget constraint for the current round.

4. The perceptual quantization incentive method for efficient federated learning in communication according to claim 1, characterized in that, In step (4), the preliminary payment rule calculation formula is as follows: ， in, This indicates that the server is on the client side. Total update quality collected during participation Indicates virtual culling client The total quality of updates that the system can collect afterward To update quality, The bid price for the client.

5. The perceptual quantization incentive method for efficient federated learning in communication according to claim 1, characterized in that, In step (4), the formula for scaling the initial payment amount using quadratic programming is as follows: ， in, For preliminary payment rules, This is the final payment rule.

6. The perceptual quantization incentive method for efficient federated learning in communication according to claim 1, characterized in that, In step (4), the budget constraints include: ， ， in, Due to budget constraints for the current round, For the final payment rules, For the selected client and its set of policies, Costs for the client.

7. The perceptual quantization incentive method for efficient federated learning in communication according to claim 1, characterized in that, The following steps are included before each of the previous steps: At the start of each round of federated learning, the server broadcasts the current global model parameters to all participating candidate clients; the server can receive parameters that the clients obtain after receiving the model and training it locally, which can characterize the local data distribution and update potential.

8. The perceptual quantization incentive method for efficient federated learning in communication according to claim 1, characterized in that, The following steps are included after each step: The server sends a selection notification and the assigned quantization strategy to the selected client; Subsequently, the following process is performed on the client: the selected client performs stochastic gradient descent model training on its local machine to obtain updated gradients, and performs quantization compression operation on the local gradient according to the quantization strategy bit width obtained in step (3), and finally uploads the compressed and quantized gradient matrix to the server.

9. A perceptual quantization excitation device for efficient federated learning in communication, characterized in that, include: The bidding receiving module receives various quantitative strategies and their corresponding bidding information submitted by each client. The quality assessment module is used to evaluate the quality of model updates for each client at different quantization levels using a quantization impact function. The Lagrange node selection module is used to replace the total budget constraint with the sum of bids from selected clients, thereby decoupling the problem. In the node selection phase, the objective function is constructed using the Lagrange duality method, and the optimal selected client and quantization strategy set are solved. The secondary planning payment scaling module determines the initial payment rules that satisfy the strict dominance strategy incentive compatibility based on the VCG mechanism, and then scales and adjusts the initial payment amount through secondary planning to meet the budget constraints, and calculates the final amount to be paid to each selected client.

10. A computer system, comprising: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the perceptual quantization incentive method for efficient federated learning for communication as described in any one of claims 1-8.

Images