Blockchain-based audit budget-constrained training authenticity verification method for federated learning

Abstract

The present application relates to a kind of audit budget restricted training authenticity verification method based on blockchain federated learning, belong to privacy computing field, this method is based on the game of training node and audit committee, the deterrent audit intensity threshold of target node is calculated, in combination with the comparison result of target audit intensity and deterrent audit intensity threshold, whether node is inclined to select dishonest behavior is judged.The audit income calculation rule after dishonest behavior is found out is clear.After in combination with dishonest probability and audit income, the expected net income of calculation is calculated, so as to determine the audit priority of each node, finally under the budget constraint, according to priority and audit cost, the audit set that needs to accept training authenticity verification is generated, training authenticity verification is executed to the node in set and the result is output.The present application converts audit under limited budget into quantifiable priority calculation problem, directly solves the verification problem whether training node is truly completed in federated learning specified local training.

Description

Technical Field

[0001] This invention relates to the field of privacy computing technology, and in particular to a method for verifying the authenticity of budget-constrained training based on blockchain federated learning. Background Technology

[0002] Existing blockchain federated learning systems typically include a task publisher, training nodes, a blockchain network, smart contracts, and an executor for model aggregation or result confirmation. The general process is as follows: (1) The task publisher publishes a federated learning task via the blockchain or smart contract, setting the training rounds, model structure, training hyperparameters, reward rules, and participation rules. (2) Training nodes register to participate in the task and download the initial global model. (3) Training nodes train the model off-chain on their local private dataset. (4) Training nodes submit local model updates, parameter summaries, or related training results. (5) The system filters, confirms, or aggregates the submitted model updates to generate a new global model. (6) The blockchain records key states, and the smart contract executes reward allocation, staking penalties, or state updates according to preset rules.

[0003] Existing blockchain-based federated learning systems typically use the blockchain as a trusted infrastructure for task publishing, model update recording, aggregation result confirmation, and incentive settlement. Since the entire model training process is difficult to execute directly on-chain, existing solutions usually maintain local training off-chain, with training nodes submitting model updates, which are then confirmed and recorded through a blockchain network or committee mechanism.

[0004] A relatively close existing technology is a blockchain federated learning scheme based on committee consensus. This type of scheme typically selects a portion of the nodes participating in training to form a committee. The committee is responsible for receiving model updates, verifying update quality, performing aggregation confirmation, and writing key results into the blockchain. A typical process includes: (1) The task publisher publishes the federated learning task in the blockchain network and sets training parameters and participation rules. (2) The training nodes complete model training locally and submit model updates to the blockchain or the committee. (3) The committee performs cross-checking, quality assessment, or consistency confirmation on the model updates submitted by the training nodes. (4) The committee filters out abnormal or low-quality updates and performs aggregation based on valid updates. (5) The aggregation results, verification records, and reward information are written into the blockchain, and subsequent settlement is completed by a smart contract. For example, BFLC-type schemes improve consensus efficiency and system scalability in blockchain federated learning through committee consensus. Their committee is used to confirm model updates and reduce the communication and consensus burden caused by all nodes participating in verification. This type of scheme demonstrates that the committee can serve as the execution and confirmation entity in blockchain federated learning.

[0005] Another type of approach is blockchain-based federated learning schemes that rely on robust aggregation or model update filtering. These schemes focus on whether model update results are anomalies, typically identifying suspicious updates based on model parameter similarity, validation set performance, gradient direction, anomaly scores, or robust aggregation rules. For example, Biscotti combines a blockchain system with privacy-preserving federated learning and introduces collaborative verification and robust aggregation concepts in the consensus phase to reduce the impact of malicious model updates on the global model. While this approach enhances the stability of aggregation results, its core judgment still primarily relies on the model update results submitted by nodes.

[0006] The technologies mentioned above are all geared towards blockchain-based federated learning scenarios, utilize blockchain to record the training interaction process, and allow model updates to be evaluated and confirmed by a committee or similar implementing body.

[0007] However, this type of technology primarily addresses the question of whether the model update results are acceptable, rather than directly resolving whether the training nodes have truly completed the required local training. If training nodes reduce the number of local training rounds, reuse historical models, or upload low-quality but numerically close-to-normal updates, relying solely on model update results for evaluation is insufficient to fully prove the authenticity of the training process. Summary of the Invention

[0008] In view of this, it is necessary to provide a method and apparatus for verifying the authenticity of audited budget-constrained training based on blockchain federated learning, so as to solve the problem that the existing technology cannot directly solve whether the training node has truly completed the prescribed local training.

[0009] To address the aforementioned problems, in a first aspect, this invention provides a method for verifying the authenticity of budget-constrained training based on blockchain federated learning, applied to committees in blockchain federated learning, comprising: The model updates submitted by the training nodes are evaluated to obtain the probability of dishonest behavior by the target training node. Based on the audit game between the training node and the committee, the deterrent audit strength threshold of the target training node is calculated. The deterrent audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. Based on the deterrence audit intensity threshold and the committee's target audit intensity for verifying the training authenticity of target training nodes, it is determined that target training nodes tend to choose dishonest behavior. Target training nodes are then audited as priority audit targets. When the dishonest behavior of a target training node is successfully detected, audit benefits are obtained based on audit rewards. Expected net income is determined based on the probability of dishonest behavior and audit benefits, and audit priority is determined based on expected net income. Under the constraint of the audit budget, an audit set that needs to undergo training authenticity verification is generated based on the audit priority and audit cost of each training node, and specific training authenticity verification is performed on the training nodes within the audit set to obtain the audit results.

[0010] In one possible implementation, calculating the deterrence audit strength threshold of the target training node based on the audit game between the training node and the committee includes: Based on the strategy of the training node and the strategy of the committee, the expected reward of the target training node when it chooses dishonest training and the expected reward of the target training node when it chooses honest training are determined. The strategy of the training node is to choose between honest training and dishonest behavior, and the strategy of the committee is to assign a vector of audit intensity to all training nodes under the audit budget constraint. When the expected return of the target training node when it chooses honest training is equal to the expected return of the target training node when it chooses dishonest training, the deterrence audit strength threshold of the target training node is determined.

[0011] In one possible implementation, determining that a target training node is prone to dishonest behavior based on a deterrence audit strength threshold and the target audit strength of the committee's training authenticity verification of the target training node includes: When the target audit strength is less than the deterrence audit strength threshold, it is determined that the target training node tends to choose dishonest behavior; When the target audit strength is equal to the deterrence audit strength threshold, the target training node is determined to be in a policy indifference state. When the target audit strength is greater than the deterrence audit strength threshold, it is determined that the target training node tends to choose honest behavior.

[0012] In one possible implementation, determining the expected net benefit based on the probability of dishonest behavior and the audit benefit, and determining the audit priority based on the expected net benefit, includes: Obtain the audit strength of the committee on the target training node, and obtain the probability that the committee can detect dishonest behavior after auditing the target training node; The joint probability is determined based on the probability of the dishonest behavior, the audit intensity, and the probability that the dishonest behavior can be detected after the audit. The expected audit benefit is determined based on the joint probability and the audit benefit. The expected net benefit of the target training node is determined based on the expected audit benefits and expected audit costs, and the audit priority is determined based on the expected net benefit.

[0013] In one possible implementation, the expression for the audit priority is: In the formula, Represents training nodes In the Audit priority for each iteration; Represents training nodes In the The probability of dishonest behavior in each iteration; Represents training nodes In the The probability of detecting anomalies found during each round of iteration; Represents training nodes Strategic violations went undetected, resulting in anticipated systemic damage; Represents training nodes In the Audit rewards for each iteration; Represents training nodes Audit costs; Represents training nodes In the Risk score of each iteration; Represents training nodes In the Historical reputation value of each iteration; Represents the weight parameters. .

[0014] In one possible implementation, the expression for the probability of the dishonest behavior is: In the formula, Represents training nodes In the The probability of dishonest behavior in each iteration; Represents training nodes In the Risk score of each iteration; Represents training nodes In the Historical reputation value of each iteration; Represents the weight parameters. .

[0015] In one possible implementation, the dynamic generation of the audit set requiring training authenticity verification, under the constraint of the audit budget and based on audit priority and audit cost, includes the following steps: Step 1: Filter out the training nodes whose audit cost in the current round is not greater than the audit budget in the current round, and obtain the filtered training nodes; Step 2: Sort the filtered training nodes in descending order of audit priority to obtain the sorted training nodes; Step 3: Determine in order whether the sorted training nodes can be added to the audit set; Step 4: If the sum of the audit budget used in the current round and the audit cost of the target training node is not greater than the audit budget, add the target training node to the audit set. The audit budget used in the next round is equal to the audit budget used in the current round plus the audit cost of the target training node in the current round, and then proceed to Step 1.

[0016] One possible implementation also includes: Based on the audit results, training nodes confirmed to have engaged in dishonest behavior will be subject to staking deduction, reward recovery, and reputation reduction.

[0017] Secondly, the present invention also provides a device for verifying the authenticity of audited budget-constrained training based on blockchain federated learning, comprising: The probability determination module for dishonest behavior is used to evaluate the model updates submitted by the training nodes and obtain the probability of dishonest behavior of the target training node. The deterrence audit strength threshold determination module is used to calculate the deterrence audit strength threshold of the target training node based on the audit game between the training node and the committee. The deterrence audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. The audit benefit determination module is used to determine the target training node's tendency to choose dishonest behavior based on the deterrent audit intensity threshold and the target audit intensity of the committee's training authenticity verification of the target training node. The target training node is then selected as the priority audit target. When the dishonest behavior of the target training node is successfully detected, audit benefits are obtained based on the audit reward. The audit prioritization module is used to determine the expected net benefit based on the probability of dishonest behavior and the audit benefit, and to determine the audit priority based on the expected net benefit. The audit result determination module is used to generate an audit set that needs to undergo training authenticity verification based on the audit priority and audit cost of each training node, under the constraint of the audit budget, and to perform specific training authenticity verification on the training nodes within the audit set to obtain the audit results.

[0018] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the audit budget-constrained training authenticity verification method based on blockchain federated learning described in any of the above implementations.

[0019] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, can implement the steps in the audit budget-constrained training authenticity verification method based on blockchain federated learning described in any of the above implementations.

[0020] The beneficial effects of this invention are as follows: The audit budget-constrained training authenticity verification method based on blockchain federated learning provided in this embodiment first evaluates the model updates submitted by the training nodes to obtain the probability of dishonest behavior by the target training node. Then, based on the audit game between the training node and the committee, it calculates the deterrent audit strength threshold for the target training node. The deterrent audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. The audit game is used to characterize the behavioral choices of the training node under the influence of training cost, cheating benefits, audit probability, and penalty strength. Furthermore, based on the deterrent audit strength threshold and the target audit strength of the committee performing training authenticity verification on the target training node, the target training node is selected as a priority audit target. When honest behavior is successfully verified, audit benefits are obtained based on audit rewards. Further, the expected net benefit is determined based on the probability of dishonest behavior and the audit benefit, and audit priorities are determined based on the expected net benefit. An audit priority index is designed so that the committee can prioritize verifying nodes with a higher probability of dishonest behavior, greater potential system losses, and higher unit audit cost benefits under limited budget constraints. Finally, under the constraint of the audit budget, an audit set for training authenticity verification is generated based on the audit priority and audit cost of each training node. Specific training authenticity verification is then performed on the training nodes within the audit set to obtain the audit results. Introducing audit budget constraints in blockchain federated learning prevents training authenticity verification from being designed to perform the same verification intensity on all nodes, instead allowing for differentiated arrangements based on node risk and budget conditions. This invention transforms auditing under limited budgets into a quantifiable priority calculation problem, directly solving the verification challenge in federated learning of whether training nodes have truly completed the prescribed local training. Attached Figure Description

[0021] Figure 1 A flowchart illustrating an embodiment of a blockchain-based federated learning method for verifying the authenticity of budget-constrained training. Figure 2 This is a structural diagram of the blockchain federated learning system of the present invention; Figure 3 This is a flowchart illustrating the verification process for training authenticity under budget constraints in this invention. Figure 4 This is a schematic diagram illustrating the audit game theory and audit priority calculation principle of this invention. Figure 5 This is a diagram illustrating the cross-round state feedback mechanism of the present invention; Figure 6 A schematic flowchart of an embodiment of a blockchain-based federated learning audit budget-constrained training authenticity verification device provided by the present invention; Figure 7 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0024] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0025] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0026] Before demonstrating the embodiments, the following terms will be explained.

[0027] Blockchain-based Federated Learning (BCFL) refers to a collaborative training paradigm that incorporates the decentralized, immutable, and distributed consensus characteristics of blockchain into federated learning. The blockchain is used to record training tasks, model updates, validation results, aggregation results, and state changes. The local training process in federated learning is typically completed independently by off-chain training nodes.

[0028] Training nodes: These are the nodes that participate in the federated learning task. Training nodes use local private data to train the model and submit model updates after training is complete.

[0029] The committee, comprised of some training nodes, is responsible for verifying and executing model updates, verifying training validity, selecting effective updates, and aggregating global models. It also writes key verification statuses and aggregation results to the blockchain.

[0030] Model update evaluation: This refers to the process by which the committee evaluates the model updates submitted by the training nodes. This process can detect model updates based on the committee's local dataset or evaluation rules, and generate a risk score for the node for the current round.

[0031] Training authenticity verification: This refers to the process of further verifying whether the training nodes have completed local training as required. This process is performed on the nodes selected for the verification set to determine whether they have engaged in dishonest behavior such as lazy training, uploading outdated models, submitting low-quality updates, or falsifying training results.

[0032] Audit budget: refers to the maximum resource constraint that the system can use for training authenticity verification in a certain round of training, including computation, communication, verification and related execution costs.

[0033] Audit intensity: refers to the probability or intensity with which the committee verifies the training authenticity of a training node in a given round. The greater the audit intensity, the higher the likelihood that a node will enter the training authenticity verification process.

[0034] Audit set: refers to the set of training nodes selected to perform training authenticity verification under the current audit budget constraints.

[0035] Dishonest behavior: refers to actions taken by training nodes to reduce their own computational or communication costs, including reducing the number of training rounds, uploading outdated models, submitting low-quality updates, or falsifying training results.

[0036] Historical reputation: refers to the state variables maintained by the system based on the node's historical training performance, model update evaluation results, and training authenticity verification results, which are used to reflect the stability of the node's past behavior.

[0037] Audit Priority: This refers to an indicator used to measure whether a node should be prioritized in the training authenticity verification process, given a limited audit budget. This indicator comprehensively considers the probability of dishonest behavior by the node, the probability of audit detection, the system losses caused by undetected dishonest behavior, the benefits of committee auditing, and the audit costs of the node.

[0038] This invention provides a method and apparatus for verifying the authenticity of budget-constrained training based on blockchain federated learning, which will be described in detail below.

[0039] Figure 1 A schematic flowchart of an embodiment of the audit budget-constrained training authenticity verification method based on blockchain federated learning provided by the present invention is shown below. Figure 1 As shown, the method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning includes: S101. Evaluate the model updates submitted by the training nodes to obtain the probability of dishonest behavior of the target training node. S102. Based on the audit game between the training node and the committee, calculate the deterrent audit strength threshold of the target training node. The deterrent audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. S103. Based on the deterrence audit intensity threshold and the committee's target audit intensity for verifying the authenticity of training on target training nodes, determine that target training nodes tend to choose dishonest behavior, and prioritize auditing target training nodes. When the dishonest behavior of target training nodes is successfully detected, obtain audit benefits based on audit rewards. S104. Determine expected net income based on the probability of dishonest behavior and audit benefits, and determine audit priority based on expected net income; S105. Under the constraint of the audit budget, an audit set that needs to undergo training authenticity verification is generated according to the audit priority and audit cost of each training node, and specific training authenticity verification is performed on the training nodes in the audit set to obtain the audit results.

[0040] Compared with existing technologies, the audit budget-constrained training authenticity verification method based on blockchain federated learning provided in this embodiment first evaluates the model updates submitted by training nodes to obtain the probability of dishonest behavior by the target training node. Then, based on the audit game between the training node and the committee, it calculates the deterrent audit strength threshold for the target training node. The deterrent audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. The audit game is used to characterize the behavioral choices of the training node under the influence of training costs, cheating benefits, audit probability, and penalty strength. Furthermore, based on the deterrent audit strength threshold and the target audit strength of the committee performing training authenticity verification on the target training node, the target training node is selected as a priority audit target for auditing. Upon successful verification of a behavior, audit revenue is awarded based on audit rewards. Further, the expected net revenue is determined based on the probability of dishonest behavior and the audit revenue. Audit priorities are then established based on this expected net revenue, and an audit priority index is designed to allow the committee to prioritize verification of nodes with a higher probability of dishonest behavior, greater potential system losses, and higher unit audit cost-benefit under limited budget constraints. Finally, under the constraint of the audit budget, an audit set requiring training authenticity verification is generated based on the audit priority and audit cost of each training node. Specific training authenticity verification is then performed on the training nodes within the audit set to obtain the audit results. This invention introduces audit budget constraints into blockchain federated learning, preventing training authenticity verification from being designed to perform the same verification intensity on all nodes. Instead, it differentiates the arrangements based on node risk and budget conditions. This invention transforms auditing under limited budgets into a quantifiable priority calculation problem, directly solving the verification challenge in federated learning of whether training nodes have truly completed the prescribed local training.

[0041] The blockchain-based federated learning audit budget-constrained training authenticity verification method provided in this embodiment can be applied to a blockchain-based federated learning audit budget-constrained training authenticity verification system. This system can be a software system running on a terminal device. The terminal device can be a tablet computer, in-vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), mobile phone, etc. This embodiment does not impose any restrictions on the specific type of terminal device.

[0042] like Figure 2 As shown, the system structure of this embodiment includes the following components: (1) Task Publisher: The task publisher is responsible for initiating blockchain federated learning tasks and publishing task information through the blockchain network or smart contracts. The task publishing information includes training tasks, initial models, training rounds, training parameters, reward rules, audit budget, and participation rules. After the training task is completed, the task publisher receives the final aggregated model.

[0043] (2) Training Nodes: Training nodes are the main participants in the training process of this invention. After registering to participate in the task, training nodes download the initial model from the blockchain network and complete off-chain training using local private data. After training is completed, training nodes submit local model updates to the blockchain network or committee. Since the training process is completed off-chain, training nodes may engage in dishonest behavior such as reducing the number of training rounds, reusing historical models, uploading low-quality updates, or falsifying training results. Therefore, training nodes are the main objects of model update evaluation and training authenticity verification in this invention.

[0044] (3) Committee: The committee, composed of some training nodes, is the verification and aggregation executor in this invention. The committee first evaluates the model updates submitted by the training nodes and generates a risk score for the current round based on the evaluation results. Subsequently, the committee calculates the audit priority by combining the node's historical reputation, detection probability, potential system loss, audit reward, and audit cost, and generates the current round audit set under audit budget constraints. For training nodes entering the audit set, the committee further performs training authenticity verification to determine whether they have truly completed the prescribed local training. After completing the evaluation and audit, the committee selects valid model updates and performs global aggregation.

[0045] (4) Blockchain Network: The blockchain network is responsible for recording key states during system operation, including task release information, model update summaries, audit sets, model update evaluation results, training authenticity verification results, aggregation results, and node state changes. The blockchain network does not directly execute the complete local training process, but instead undertakes the functions of recording, confirming, and tracing to ensure the credibility and auditability of key interaction processes.

[0046] (5) Smart Contracts: Smart contracts are responsible for executing incentive distribution, staking penalties, reward recovery, and reputation status synchronization based on the verification and aggregation results submitted by the committee. When a node passes the training authenticity verification, the smart contract retains its staking and updates its reputation according to the rules; when a node is confirmed to have engaged in dishonest behavior, the smart contract executes staking deduction, reward recovery, and reputation reduction; for nodes that are not included in the audit set, the smart contract updates their reputation status based on the model update evaluation results.

[0047] Through the above structure, the present invention combines off-chain local training, on-chain state recording, committee verification and execution, and smart contract incentive settlement. Figure 2 In this architecture, task publishers release tasks to the blockchain, training nodes download the initial model and upload local model updates, a committee performs model update evaluation, training authenticity verification, and model aggregation, and the aggregation results are written back to the chain. Smart contracts then distribute incentives and update the state based on the verification results. This structure enables the system to selectively audit key nodes under budget constraints without performing full training authenticity verification on all training nodes.

[0048] It should be noted that, in order to better illustrate the method flow of this embodiment, the key parameters of this embodiment are listed below. Training node set The probability that node i will choose dishonest behavior in round t. Audit strength for node i Round t General Audit Budget Audit cost of node i The probability of a node being detected as an anomaly during auditing. The expected system loss when node dishonest behavior goes undetected. The node's risk score in the current round. Node historical reputation value Total penalty after a node is confirmed to be dishonest. And the priority indicators of the unit's audit budget. .

[0049] Among them, risk score From model update evaluation. Reputation value. Status feedback from previous rounds. It is used to measure the expected returns when a unit of audit budget is allocated to a certain node.

[0050] like Figure 3 As shown, the present invention operates according to the following steps: Step 1: Task Issuance and Initialization The task publisher releases the federated learning task to the blockchain network via a smart contract, setting the training rounds, model structure, training hyperparameters, task rewards, and the maximum available audit budget for the current round. Training nodes register to participate in the task by staking tokens and download the initial global model.

[0051] Step 2: Off-chain local training and model update submission Training nodes use local private data to train the model. After training is complete, the training nodes submit model updates to the system. The blockchain network records a summary of the model updates or related state information.

[0052] Step 3: Model Update Evaluation and Current Round Risk Score Generation The committee performs model update evaluation on all model updates submitted by training nodes. Model update evaluation is a lightweight outcome-based evaluation used to initially determine whether there are any abnormal characteristics in the model updates submitted by the nodes, and to provide risk input for the current round of audit priority calculation.

[0053] Let the set of training nodes in round t be: , For any training node The model update it submits is denoted as The committee evaluates the models submitted by nodes based on the local test set. For ease of implementation, this invention normalizes the model update evaluation results to the current round's risk score. ,in, The larger the value, the more suspicious the model update submitted by node i in round t is. The smaller the value, the closer the current update of the node is to a normal update.

[0054] One specific implementation method is that the committee calculates node risk from the following three aspects: , in And satisfy .

[0055] This indicates the degree of anomaly in the normalized loss of node i on the committee evaluation dataset. If the node update leads to a significant increase in the validation loss, then... Relatively large.

[0056] This indicates the degree of anomaly in the normalized distance between the model update of node i and the statistical center of the current round of normal updates. If the node update differs significantly from the updates of most nodes, then... Relatively large.

[0057] This represents the normalized similarity between the model update direction of node i and the reference update direction. If the node's update direction is similar to the normal training direction, then... Larger; if the directional difference is large, then 1 Relatively large.

[0058] Therefore, the current round risk score It comprehensively reflects the degree of suspicion of node updates in terms of verification loss, update magnitude, and update direction.

[0059] Step 4: Training the audit game model between nodes and the committee The game theory model in this embodiment refers to an audit game between the training node and the committee surrounding the verification of training authenticity. The training node, as the auditee, chooses between honest training and dishonest behavior. The committee, as the audit executor, decides, within a limited audit budget, whether and to what extent to perform training authenticity verification on the training node. This game theory model characterizes the behavioral choices of the training node under the influence of training costs, cheating benefits, audit probability, and penalty intensity, as well as the committee's audit decisions under the influence of audit costs, detection probability, potential system losses, and audit rewards. Its core function is to explain how the system increases the expected cost of dishonest behavior by the training node through a limited audit budget, penalty mechanisms, and reputation feedback, and to provide a basis for audit priority calculation and audit set generation.

[0060] After completing the model update evaluation, the committee needs to determine whether to perform further training authenticity verification on the nodes. Since training authenticity verification consumes additional audit budget, the system cannot perform training authenticity verification indiscriminately on all nodes. Therefore, this invention models the relationship between training nodes and the committee as an audit game.

[0061] like Figure 4 As shown, for any training node i, its policy set is defined as: Here, H represents honest training, and C represents dishonest behavior. Dishonest behavior includes reducing the number of training epochs, uploading outdated models, submitting low-quality updates, or fabricating training results.

[0062] The probability that node i chooses dishonest behavior in round t is denoted as: .

[0063] The committee's strategy is represented as a vector that assigns audit strengths to all training nodes: .

[0064] in, This represents the probability that the committee will perform training authenticity verification on node i in round t.

[0065] like This indicates that the node only accepts model update evaluation and does not enter the training authenticity verification process. The larger the value, the higher the likelihood that node i will be subject to further auditing.

[0066] Considering node audit costs and the total audit budget, the committee's feasible strategy space is defined as follows: in, This represents the audit cost required to perform training authenticity verification on node i. Let t represent the total audit budget for round t. This formula indicates that the committee can allocate different audit intensities to different nodes, but the sum of the expected audit budgets consumed by all nodes cannot exceed the total budget for the current round.

[0067] The expected return of node i when it chooses honest training in round t is: , in, This represents the base reward that node i can obtain after completing training in round t. This represents the cost required for a node to honestly complete local training.

[0068] If node i chooses dishonest training, its expected return is: in, This indicates that the cost incurred by a node when it engages in dishonest behavior is typically lower than the cost of training it to be honest. This represents the additional opportunity gain a node obtains through dishonest behavior; This indicates the probability that dishonest behavior will be discovered after a node is audited; This represents the total penalty a node incurs after being found to have engaged in dishonest behavior. Defined as: .

[0069] in, This indicates the remaining staked amount of node i in round t. Indicates the percentage of penalties for pledged shares. This indicates that the task reward has been recovered.

[0070] From formulas (2) and (3), it can be seen that whether a node chooses dishonest behavior depends on whether the benefits of dishonest behavior outweigh the benefits of honest training. According to the definition of balance, when At this time, the node benefits equally from honest training and dishonest behavior, and it is impossible for the training node to obtain better utility by changing its strategy. Substituting formulas (2) and (3) into the equation, we can obtain: remember To obtain the deterrence audit strength threshold for node i after processing, we have: In some embodiments of the present invention, calculating the deterrent audit strength threshold of the target training node based on the audit game between the training node and the committee includes: Based on the strategy of the training node and the strategy of the committee, the expected reward of the target training node when it chooses dishonest training and the expected reward of the target training node when it chooses honest training are determined. The strategy of the training node is to choose between honest training and dishonest behavior, and the strategy of the committee is to assign a vector of audit intensity to all training nodes under the audit budget constraint. When the expected return of the target training node when it chooses honest training is equal to the expected return of the target training node when it chooses dishonest training, the deterrence audit strength threshold of the target training node is determined.

[0071] In some embodiments of the present invention, determining that a target training node tends to choose dishonest behavior based on a deterrence audit strength threshold and the target audit strength of the committee's training authenticity verification of the target training node includes: When the target audit strength is less than the deterrence audit strength threshold, it is determined that the target training node tends to choose dishonest behavior; When the target audit strength is equal to the deterrence audit strength threshold, the target training node is determined to be in a policy indifference state. When the target audit strength is greater than the deterrence audit strength threshold, it is determined that the target training node tends to choose honest behavior.

[0072] in, This represents the minimum audit strength required to prevent a node from gaining higher expected returns through dishonest behavior. This threshold is determined by the training cost difference, the dishonest opportunity gain, the detection probability, and the penalty strength.

[0073] when This indicates that the audit intensity allocated to node i by the committee is insufficient, the expected reward for dishonest behavior is still higher than the reward for honest training, and the node is more inclined to choose dishonest behavior; when At this time, the expected rewards of choosing honest training and dishonest behavior are equal, and the node is in a policy indifference state. when This indicates that the audit intensity allocated to node i by the committee has exceeded the deterrent threshold. The expected loss after the dishonest behavior is discovered is higher than the cost savings and opportunity gains. Therefore, it is difficult for the node to obtain higher expected benefits through dishonest behavior.

[0074] so, It is the key threshold for node behavior constraints. This is the actual audit intensity allocated by the committee. The comparison between the two determines whether a node has a reward incentive to choose dishonest behavior in the current round.

[0075] Given a limited audit budget, the committee cannot rely solely on... Perform adequate audits on all nodes. If the budget allows, the committee can ensure that the audit intensity for each node meets the required standards. This makes it difficult for any node to gain higher rewards through dishonest behavior. If the budget is insufficient, the committee needs to prioritize auditing nodes that are more worthy of auditing.

[0076] If node i exhibits dishonest behavior that goes undetected, its model update may enter the aggregation process and cause system losses, denoted as . If node i is audited and found to have engaged in dishonest behavior, the committee can receive an audit reward from the forfeited stake: in This represents the percentage of the reward the committee receives from the forfeitured collateral. Therefore, if node i's dishonest behavior is successfully verified, the committee's reward is: , In round t, the probability that node i chooses dishonest behavior is The audit strength of the committee's allocation to node i is The probability of detecting dishonest behavior after auditing is... Therefore, the joint probability that node i chooses dishonest behavior, is audited by the committee, and is successfully identified is: , Therefore, the expected audit benefit obtained by the committee at node i is: , At the same time, performing training authenticity verification on node i incurs audit costs. Because the committee's audit strength for node i is Therefore, its expected audit cost is: , In summary, the committee's expected net benefit at node i equals the expected audit benefit minus the expected audit cost, i.e.: After obtaining the optimal response from the training nodes, the committee's goal is to allocate appropriate audit intensity to different nodes within the constraints of the audit budget in order to maximize overall regulatory benefits.

[0077] Step 5: Audit Priority Calculation From formula (5), we can see that the committee's total expected return for all training nodes in round t is: In round t, the committee first represents the probability of dishonest behavior of each training node as a vector: in, Let represent the probability that node i chooses dishonest behavior in round t. Meanwhile, the committee will represent the audit intensity assigned to each node as a vector: in, This represents the probability or strength of the committee performing training authenticity verification on node i. It is the probability vector of dishonest behavior at a given node. Under these conditions, the committee needs to consider the feasible audit strategy space. Select an appropriate audit strength vector. This will increase the total expected audit revenue for round t. To maximize. Therefore, the committee's optimization objective can be expressed as: This formula indicates that, under audit budget constraints, the committee determines how to allocate audit intensity to different nodes based on the probability of dishonest behavior, audit costs, detection probability, potential system losses, and audit rewards, thereby maximizing the overall net audit benefit for the current round.

[0078] To analyze the benefits of the unit's audit budget, the audit budget allocated to node i is denoted as: Then we have: Substitute into the committee's total revenue function We can obtain: Therefore, it can be seen that the audit revenue generated by node i for each unit of audit budget consumed is mainly due to Therefore, the audit priority of node i in round t is defined as follows: ,have The larger the value, the higher the governance benefit of allocating the unit audit budget to that node, and the higher the priority that node should be given to entering the training authenticity verification process. When the budget is insufficient, the committee will... Differentiated verification resources are allocated to each node. The larger the value, the higher the audit intensity the node receives in the current round.

[0079] In some embodiments of the present invention, determining the expected net benefit based on the probability of dishonest behavior and audit benefits, and determining the audit priority based on the expected net benefit, includes: Obtain the audit strength of the committee on the target training node, and obtain the probability that the committee can detect dishonest behavior after auditing the target training node; The joint probability is determined based on the probability of the dishonest behavior, the audit intensity, and the probability that the dishonest behavior can be detected after the audit. The expected audit benefit is determined based on the joint probability and the audit benefit. The expected net benefit of the target training node is determined based on the expected audit benefits and expected audit costs, and the audit priority is determined based on the expected net benefit.

[0080] Step 6: Based on model updates, evaluate and approximate the probability of dishonest behavior using historical reputation.

[0081] The above audit priorities Includes the probability of dishonest behavior of nodes However, in actual operation, the committee cannot directly observe whether a node intends to act dishonestly. Therefore, it needs to combine the current round of model update evaluation results with historical reputation status to assess the situation. Make an approximate estimate.

[0082] Let the historical reputation of node i in round t be . . The larger the value, the more stable the node's historical behavior; The smaller the value, the more unstable the node's historical behavior.

[0083] The committee estimates the probability of dishonest behavior based on the current round of model updates and historical reputation status. In some embodiments of the invention, the expression for the probability of dishonest behavior is: In the formula, Represents training nodes In the The probability of dishonest behavior in each iteration; Represents training nodes In the Risk score of each iteration; Represents training nodes In the Historical reputation value of each iteration; Represents the weight parameters. This is used to adjust the impact of the current round of model update evaluation results and historical reputation status on the estimation of the probability of dishonest behavior. When A larger value indicates that the current round of model update for the node is questionable. It increases accordingly. When A smaller value indicates a node with poor historical reputation. Larger This increases accordingly. When a node has low risk in the current round and high historical reputation, A smaller value indicates a lower probability that the node will be identified as a high-risk object in the current round. Therefore, the audit priority of node i in round t can be written as: In the formula, Represents training nodes In the Audit priority for each iteration; Represents training nodes In the The probability of dishonest behavior in each iteration; Represents training nodes In the The probability of detecting anomalies found during each round of iteration; Represents training nodes Strategic violations went undetected, resulting in anticipated systemic damage; Represents training nodes In the Audit rewards for each iteration; Represents training nodes Audit costs; Represents training nodes In the Risk score of each iteration; Represents training nodes In the Historical reputation value of each iteration; Represents the weight parameters. .

[0084] Formula (9) illustrates that the audit priority of a node is determined by five factors: current round risk score, historical reputation status, audit detection probability, potential system loss and audit reward, and audit cost. The higher the current round risk, the lower the historical reputation, the higher the detection probability, the greater the potential loss and audit reward, and the lower the audit cost, the higher the node audit priority.

[0085] Step 7: Generate Audit Set Based on the audit priorities of each node, the committee generates the current round audit set under the budget constraints of the t-th round audit. The audit set must satisfy the following: in, This represents the set of nodes selected in round t to perform training authenticity verification. This represents the total audit budget for round t.

[0086] One specific implementation method is as follows: First, the committee calculates the current round risk score for all training nodes. and read the node's historical reputation Secondly, the committee estimates the probability of dishonest behavior by nodes according to formula (8). Then, the committee calculates the node audit priority according to formula (9). Subsequently, the committee follows... The training nodes are sorted from highest to lowest, and each node is then checked to see if it can be added to the audit set. Let the audit budget used in round t be . Initialized as: , If for a certain node i: Then add node i to the audit set. and update the audit budget already used: If the remaining budget is insufficient to support adding the next node, stop the selection process and output the current round of audit set. For nodes that are included in the audit set, the committee performs training authenticity verification; for nodes that are not included in the audit set, the system only updates their status based on the model update evaluation results and historical reputation status.

[0087] Thus, steps three through seven form a complete logical chain: Model update evaluation → , → → → The system first obtains the current round's risk score through model updates and assessments. Then, it combines historical reputation data to estimate the probability of dishonest behavior by nodes. Further, it calculates the audit priority within the unit audit budget and finally generates a training authenticity verification set under the audit budget constraint. This process ensures that the limited audit budget is prioritized for nodes that require more verification, rather than performing indiscriminate audits on all training nodes.

[0088] In some embodiments of the present invention, the step of dynamically generating an audit set that needs to undergo training authenticity verification based on audit priority and audit cost under the constraint of audit budget includes the following steps: Step 1: Filter out the training nodes whose audit cost in the current round is not greater than the audit budget in the current round, and obtain the filtered training nodes; Step 2: Sort the filtered training nodes in descending order of audit priority to obtain the sorted training nodes; Step 3: Determine in order whether the sorted training nodes can be added to the audit set; Step 4: If the sum of the audit budget used in the current round and the audit cost of the target training node is not greater than the audit budget, add the target training node to the audit set. The audit budget used in the next round is equal to the audit budget used in the current round plus the audit cost of the target training node in the current round, and then proceed to Step 1.

[0089] Step 8: Effectively update filters and global aggregation The committee selects valid model updates based on the model update evaluation results and training authenticity verification results. For model updates that pass evaluation and verification, the committee uses them for global model aggregation. Model updates from nodes confirmed to have engaged in dishonest behavior are not included in the normal aggregation process. After aggregation is complete, the committee writes the aggregation results and key verification status to the blockchain network.

[0090] In some embodiments of the present invention, it further includes: Based on the audit results, training nodes confirmed to have engaged in dishonest behavior will be subject to staking deduction, reward recovery, and reputation reduction.

[0091] Step 9: Status Feedback and Incentive Update After the current round of training authenticity verification is completed, the system updates the node's staking status and reputation status based on the verification results, and passes the updated status to subsequent training rounds.

[0092] like Figure 5 As shown, the first In this round, the system first generates a risk score through model updates and assessments, then generates an audit set through budget-constrained auditing, and performs training authenticity verification on the audit set. The results of this round are divided into three categories: (1) Nodes that have been audited and confirmed to have engaged in dishonest behavior. For such nodes, the system will implement staking deduction, reward recovery, and reputation reduction.

[0093] (2) Audited and approved nodes. For such nodes, the system maintains their pledged status and updates their reputation status based on the audit results.

[0094] (3) Nodes that have not entered the audit set. For such nodes, the system adjusts their reputation status based on the current round model update evaluation results.

[0095] After the above processing, the node status is updated to the next round's status, including the staking status: and credit status The updated node status serves as input for the audit priority calculation and audit set generation in round t+1. In the next round, the system estimates the following based on the new risk score and updated reputation status: And further calculate the priority of the next round of audits: The next audit set is then generated: As a result, the current round of evaluation results, audit results, pledge status and credit status are fed back to subsequent rounds, forming a dynamic feedback mechanism across rounds.

[0096] Staking Update Formula: Scenario 1: Nodes where dishonest behavior has been audited and confirmed: If node If the training authenticity verification results indicate that a node has engaged in dishonest behavior, the system will deduct its stake according to a preset ratio: in, This represents the penalty percentage for staking. The formula indicates the percentage of staking penalties that a node incurs after being confirmed to have engaged in dishonest behavior. Forfeited, remaining pledged amount updated to Meanwhile, the forfeited pledges can be used for audit rewards or the system reward pool. If the committee receives a percentage from the forfeited pledges... The audit reward received by the committee would be: The second scenario: Nodes that have been audited and passed the training authenticity verification. If node If the training authenticity verification results show that the node passes the audit, it means that the system has not detected any dishonest behavior. In this case, the node's stake will not be reduced. The formula indicates that nodes that have undergone authenticity verification through training maintain their original staking status.

[0097] The third scenario: Nodes that have not been included in the audit set. If node If the node has not undergone training authenticity verification, the system cannot directly confirm whether it has engaged in dishonest behavior. Therefore, no penalty will be imposed on it in the current round. Reputation update formula: Let the reputation value of node i at the end of round t be... The reputation value at the start of round t is And satisfy: .

[0098] Let the reputation of a node in round t be determined by the parameter It indicates. Among them. This indicates the cumulative strength that a node has historically been considered to have. This represents the cumulative strength of a node that has historically been considered untrustworthy. Then the node... In the The reputation value of a wheel is defined as: To avoid assigning excessively high or low reputation to new nodes due to a lack of historical behavior records, this paper adopts a neutral initialization method. , .

[0099] Scenario 1: Reputation update of audited nodes If node Let represent that node i enters the audit set in round t and undergoes training authenticity verification. Let its audit result be... ,in This indicates that the audit has been passed. This indicates that the node has been identified as cheating or non-compliant. For audited nodes, the system updates their trusted cumulative strength and untrusted cumulative strength based on the training authenticity verification results: in, The time decay factor, This is the weighted penalty coefficient. As a node passes the training authenticity verification, its compliance record is positively accumulated.

[0100] When a node passes the training authenticity verification... ,but: When a node is confirmed to have engaged in dishonest behavior ,but: Once a node is identified as cheating, its abnormal records will be increased with higher weight. Therefore, the negative impact of a single instance of cheating will not be quickly offset in the short term, but will continue to affect subsequent rounds.

[0101] The second scenario: Reputation soft updates that have not entered the audit aggregation node. In a real-world system, not all nodes undergo training authenticity verification in every round. For unaudited nodes, the system lacks direct training authenticity verification results and cannot directly classify them as honest or dishonest. In this case, this embodiment performs a soft update on their reputation based on the model update evaluation results: Wherein, the evaluation score of node i in the model update of round t is , This is a soft update weight.

[0102] Reputation score calculation and cross-wheel feedback After updating the above parameters, the reputation value of node i in round t+1 is calculated as follows: The updated reputation score will be used as input for estimating the probability of dishonest behavior in the next round: In summary, the technical effects of this embodiment include the following: First, the audit budget-constrained training authenticity verification mechanism proposed in this embodiment addresses the difficulty of directly verifying the authenticity of off-chain local training in blockchain federated learning by introducing audit budget constraints. This prevents the system from indiscriminately performing training authenticity verification on all training nodes, instead selecting nodes that require verification under a limited audit budget. This mechanism transforms training authenticity verification from full-node verification to selective verification under budget constraints, thereby constraining dishonest behavior of training nodes and reducing the system overhead of continuous verification under limited computing, communication, and verification resources.

[0103] Second, this embodiment proposes a hierarchical verification method that combines model update evaluation with training authenticity verification. This embodiment divides the verification process into two levels: model update evaluation and training authenticity verification. Model update evaluation is used to perform a lightweight screening of model updates submitted by all training nodes and generate a risk score for the current round; training authenticity verification is used to perform further verification on key nodes in the audit set to determine whether they have truly completed the required local training. This design avoids relying entirely on model update results to judge training authenticity and also avoids continuously performing high-cost audits on all nodes, thus forming a hierarchical verification mechanism of lightweight evaluation + key audits.

[0104] Third, this embodiment proposes an audit priority calculation method based on audit game theory, modeling the interaction between training nodes and the committee as an audit game. Training nodes choose between honest training and dishonest behavior, while the committee decides to allocate audit intensity to different nodes under audit budget constraints. Based on this, the invention comprehensively considers the probability of dishonest behavior by nodes, the probability of audit detection, the system loss caused by undetected dishonest behavior, the committee's audit reward, and the node audit cost to construct an audit priority index. This index measures the expected return when a unit of audit budget is invested in a particular node, enabling the committee to prioritize nodes that are more worthy of verification.

[0105] This innovation is the core of this embodiment: it transforms the question of "who to audit" under a limited budget into a quantifiable prioritization problem. This metric measures the necessity of a node being prioritized for verification in the current round.

[0106] Fourth, this embodiment protects a node behavior constraint method based on a deterrence audit strength threshold. This method calculates the deterrence audit strength threshold for a node based on honest training cost, dishonest behavior cost, opportunity gain, detection probability, and penalty intensity.

[0107] Fifth, this embodiment protects the method for approximating the probability of dishonest behavior based on the current round of model update evaluation results and historical reputation status: This method uses both the current round of risk and historical behavior status for audit priority calculation.

[0108] Sixth, this embodiment protects a method for generating audit sets under audit budget constraints. This method, based on audit priority and node audit cost, generates audit sets while satisfying... The set of objects for verifying the authenticity of the current round of training is generated under the given conditions.

[0109] To better implement the audit budget-constrained training authenticity verification method based on blockchain federated learning in this embodiment of the invention, correspondingly, as follows: Figure 6 As shown, this embodiment of the invention also provides a blockchain-based federated learning audit budget-constrained training authenticity verification device. The blockchain-based federated learning audit budget-constrained training authenticity verification device 600 includes: The probability determination module 601 for dishonest behavior is used to evaluate the model updates submitted by the training nodes and obtain the probability of dishonest behavior of the target training node. The deterrence audit strength threshold determination module 602 is used to calculate the deterrence audit strength threshold of the target training node based on the audit game between the training node and the committee. The deterrence audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. The audit benefit determination module 603 is used to determine the target training node's tendency to choose dishonest behavior based on the deterrent audit intensity threshold and the target audit intensity of the committee's training authenticity verification of the target training node, and to select the target training node as the priority audit object for auditing. When the dishonest behavior of the target training node is successfully detected, audit benefits are obtained based on audit rewards. Audit prioritization module 604 is used to determine expected net income based on the probability of dishonest behavior and audit benefits, and to determine audit priority based on expected net income. The audit result determination module 605 is used to generate an audit set that needs to undergo training authenticity verification based on the audit priority and audit cost of each training node under the constraint of the audit budget, and to perform specific training authenticity verification on the training nodes in the audit set to obtain the audit result.

[0110] The above embodiment provides a blockchain-based federated learning audit budget-constrained training authenticity verification device 600, which can implement the technical solutions described in the above embodiment of the blockchain-based federated learning audit budget-constrained training authenticity verification method. The specific implementation principles of each module or unit can be found in the corresponding content of the above embodiment of the blockchain-based federated learning audit budget-constrained training authenticity verification method, which will not be repeated here.

[0111] like Figure 7 As shown, the present invention also provides an electronic device 700. The electronic device 700 includes a processor 701, a memory 702, and a display 703. Figure 7 Only some components of the electronic device 700 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0112] In some embodiments, processor 701 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 702 or process data, such as the audit budget-constrained training authenticity verification method based on blockchain federated learning in this invention.

[0113] In some embodiments, processor 701 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 701 may be local or remote. In some embodiments, processor 701 may be implemented on a cloud platform. In some embodiments, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-cloud, or any combination thereof.

[0114] In some embodiments, memory 702 may be an internal storage unit of electronic device 700, such as a hard disk or memory of electronic device 700. In other embodiments, memory 702 may also be an external storage device of electronic device 700, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 700.

[0115] Furthermore, the memory 702 may include both internal storage units of the electronic device 700 and external storage devices. The memory 702 is used to store application software and various types of data installed on the electronic device 700.

[0116] In some embodiments, display 703 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 703 is used to display information from electronic device 700 and to display a visual user interface. Components 701-703 of electronic device 700 communicate with each other via a system bus.

[0117] In one embodiment, when processor 701 executes a blockchain-based federated learning audit budget-constrained training authenticity verification procedure stored in memory 702, the following steps can be implemented: The model updates submitted by the training nodes are evaluated to obtain the probability of dishonest behavior by the target training node. Based on the audit game between the training node and the committee, the deterrent audit strength threshold of the target training node is calculated. The deterrent audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. Based on the deterrence audit intensity threshold and the committee's target audit intensity for verifying the training authenticity of target training nodes, it is determined that target training nodes tend to choose dishonest behavior. Target training nodes are then audited as priority audit targets. When the dishonest behavior of a target training node is successfully detected, audit benefits are obtained based on audit rewards. Expected net income is determined based on the probability of dishonest behavior and audit benefits, and audit priority is determined based on expected net income. Under the constraint of the audit budget, an audit set that needs to undergo training authenticity verification is generated based on the audit priority and audit cost of each training node, and specific training authenticity verification is performed on the training nodes within the audit set to obtain the audit results.

[0118] It should be understood that when the processor 701 executes a blockchain-based federated learning audit budget-constrained training authenticity verification program in the memory 702, in addition to the functions mentioned above, it can also perform other functions, as detailed in the description of the corresponding method embodiments above.

[0119] Furthermore, the embodiments of the present invention do not specifically limit the type of electronic device 700 mentioned. Electronic device 700 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, electronic device 700 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0120] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0121] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for verifying the authenticity of budget-constrained training based on blockchain federated learning, applied to committees in blockchain federated learning, characterized in that... include: The model updates submitted by the training nodes are evaluated to obtain the probability of dishonest behavior by the target training node. Based on the audit game between the training node and the committee, the deterrent audit strength threshold of the target training node is calculated. The deterrent audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. Based on the deterrence audit intensity threshold and the committee's target audit intensity for verifying the training authenticity of target training nodes, it is determined that target training nodes tend to choose dishonest behavior. Target training nodes are then audited as priority audit targets. When the dishonest behavior of a target training node is successfully detected, audit benefits are obtained based on audit rewards. Expected net income is determined based on the probability of dishonest behavior and audit benefits, and audit priority is determined based on expected net income. Under the constraint of the audit budget, an audit set that needs to undergo training authenticity verification is generated based on the audit priority and audit cost of each training node, and specific training authenticity verification is performed on the training nodes within the audit set to obtain the audit results.

2. The method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning according to claim 1, characterized in that, The calculation of the deterrent audit strength threshold of the target training node based on the audit game between the training node and the committee includes: Based on the strategy of the training node and the strategy of the committee, the expected reward of the target training node when it chooses dishonest training and the expected reward of the target training node when it chooses honest training are determined. The strategy of the training node is to choose between honest training and dishonest behavior, and the strategy of the committee is to assign a vector of audit intensity to all training nodes under the audit budget constraint. When the expected return of the target training node when it chooses honest training is equal to the expected return of the target training node when it chooses dishonest training, the deterrence audit strength threshold of the target training node is determined.

3. The method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning according to claim 1, characterized in that, The determination of whether a target training node is prone to dishonest behavior, based on the deterrence audit strength threshold and the target audit strength performed by the committee on the training authenticity verification of the target training node, includes: When the target audit strength is less than the deterrence audit strength threshold, it is determined that the target training node tends to choose dishonest behavior; When the target audit strength is equal to the deterrence audit strength threshold, the target training node is determined to be in a policy indifference state. When the target audit strength is greater than the deterrence audit strength threshold, it is determined that the target training node tends to choose honest behavior.

4. The method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning according to claim 1, characterized in that, The determination of expected net income based on the probability of dishonest behavior and audit benefits, and the determination of audit priority based on expected net income, include: Obtain the audit strength of the committee on the target training node, and obtain the probability that the committee can detect dishonest behavior after auditing the target training node; The joint probability is determined based on the probability of the dishonest behavior, the audit intensity, and the probability that the dishonest behavior can be detected after the audit. The expected audit benefit is determined based on the joint probability and the audit benefit. The expected net benefit of the target training node is determined based on the expected audit benefits and expected audit costs, and the audit priority is determined based on the expected net benefit.

5. The method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning according to claim 1, characterized in that, The expression for the audit priority is: In the formula, Represents training nodes In the Audit priority for each iteration; Represents training nodes In the The probability of dishonest behavior in each iteration; Represents training nodes In the The probability of detecting anomalies found during each round of iteration; Represents training nodes Strategic violations went undetected, resulting in anticipated systemic damage; Represents training nodes In the Audit rewards for each iteration; Represents training nodes Audit costs; Represents training nodes In the Risk score of each iteration; Represents training nodes In the Historical reputation value of each iteration; Represents the weight parameters. .

6. The method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning according to claim 1, characterized in that, The expression for the probability of the dishonest behavior is: In the formula, Represents training nodes In the The probability of dishonest behavior in each iteration; Represents training nodes In the Risk score of each iteration; Represents training nodes In the Historical reputation value of each iteration; Represents the weight parameters. .

7. The method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning according to claim 1, characterized in that, The process of dynamically generating a set of audits requiring training authenticity verification based on audit priorities and costs, under the constraint of the audit budget, includes the following steps: Step 1: Filter out the training nodes whose audit cost in the current round is not greater than the audit budget in the current round, and obtain the filtered training nodes; Step 2: Sort the filtered training nodes in descending order of audit priority to obtain the sorted training nodes; Step 3: Determine in order whether the sorted training nodes can be added to the audit set; Step 4: If the sum of the audit budget used in the current round and the audit cost of the target training node is not greater than the audit budget, add the target training node to the audit set. The audit budget used in the next round is equal to the audit budget used in the current round plus the audit cost of the target training node in the current round, and then proceed to Step 1.

8. The method for verifying the authenticity of audited budget-constrained training based on blockchain federated learning according to claim 1, characterized in that, Also includes: Based on the audit results, training nodes confirmed to have engaged in dishonest behavior will be subject to staking deduction, reward recovery, and reputation reduction.

9. A device for verifying the authenticity of budget-constrained training based on blockchain federated learning, characterized in that, include: The probability determination module for dishonest behavior is used to evaluate the model updates submitted by the training nodes and obtain the probability of dishonest behavior of the target training node. The deterrence audit strength threshold determination module is used to calculate the deterrence audit strength threshold of the target training node based on the audit game between the training node and the committee. The deterrence audit strength threshold is the minimum audit strength required to make the expected benefits of the training node choosing honest training and dishonest behavior equal. The audit benefit determination module is used to determine the target training node's tendency to choose dishonest behavior based on the deterrent audit intensity threshold and the target audit intensity of the committee's training authenticity verification of the target training node. The target training node is then selected as the priority audit target. When the dishonest behavior of the target training node is successfully detected, audit benefits are obtained based on the audit reward. The audit prioritization module is used to determine the expected net benefit based on the probability of dishonest behavior and the audit benefit, and to determine the audit priority based on the expected net benefit. The audit result determination module is used to generate an audit set that needs to undergo training authenticity verification based on the audit priority and audit cost of each training node, under the constraint of the audit budget, and to perform specific training authenticity verification on the training nodes within the audit set to obtain the audit results.

10. An electronic device, characterized in that, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the audit budget-constrained training authenticity verification method based on blockchain federated learning as described in any one of claims 1 to 8.

Images