A noise data processing method and device based on federated learning and a medium

Abstract

The application discloses a noise data processing method and device based on federated learning, equipment and medium, relates to the fields of data processing, data privacy protection and machine learning optimization, and is used for solving the problem that noise intensity cannot be dynamically adjusted according to task characteristics in the model training process. The method comprises the following steps: comparing the privacy consumption value of a global large model with a preset privacy budget; when the privacy consumption value exceeds the preset privacy budget, stopping updating the global large model; finding the task characteristics of a training task and the noise intensity interval of the task characteristics; according to the task characteristic priority of the training task, extracting a target noise intensity interval from the noise intensity interval; and adjusting the added noise of the training task according to the target noise intensity interval. The strength of privacy protection is evaluated by the privacy budget, and the noise intensity is dynamically adjusted according to the task characteristics, so that the privacy of data in the training process is fully guaranteed, and sensitive information leakage is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing and the field of data privacy protection and machine learning optimization, in particular to a noise data processing method based on federated learning, equipment and medium. BACKGROUND

[0002] With the gradual strictness of data privacy regulations (such as GDPR, etc.), various relevant fields in our life contain a large amount of personal sensitive information, and in terms of privacy protection, the current centralized data processing and machine learning method is facing increasing compliance pressure and privacy leakage risk.

[0003] Federated learning, as a new distributed machine learning method, has attracted widespread attention. The core idea of federated learning is to train the model locally by each client and upload only the model parameters rather than the original data, thereby realizing data privacy protection and avoiding the data leakage problem in the current centralized learning method. However, federated learning still faces the problem of how to balance privacy protection and model performance in actual application, especially in the case of different task characteristics, so differential privacy technology is usually combined with federated learning technology to protect data privacy, but it cannot dynamically adjust the noise intensity according to the specific characteristics of the task in the model training, thereby affecting the efficiency and accuracy of the training. SUMMARY

[0004] The embodiments of the present application provide a noise data processing method based on federated learning, equipment and medium, which is used to solve the problem that the noise intensity cannot be dynamically adjusted according to the task characteristics in the model training process.

[0005] The embodiments of the present application adopt the following technical solutions:

[0006] On the one hand, the embodiments of the present application provide a noise data processing method based on federated learning, which comprises: comparing the privacy consumption value of a global large model with a preset privacy budget; stopping updating the global large model when the privacy consumption value exceeds the preset privacy budget; finding the task characteristics of a training task and the noise intensity interval of the task characteristics; each task characteristic has a corresponding noise intensity interval; according to the task characteristic priority of the training task, extracting a target noise intensity interval from the noise intensity interval; and adjusting the added noise of the training task according to the target noise intensity interval.

[0007] In one example, the target noise intensity interval is extracted from the noise intensity interval according to the task characteristic priority of the training task, specifically including: when there are multiple task characteristics, finding overlapping intervals between the noise intensity interval of the highest priority task characteristic and the noise intensity intervals of other task characteristics, obtaining an overlapping interval set, and extracting the target noise intensity interval from the noise intensity interval of the highest priority task characteristic according to the overlapping interval set.

[0008] In one example, the target noise intensity interval is extracted from the noise intensity interval of the highest priority task characteristic according to the overlapping interval set, specifically including: when the overlapping interval set is empty, determining the noise intensity interval of the highest priority task characteristic as the target noise interval.

[0009] In one example, the method further includes: when the overlapping interval set is not empty, determining whether there is a task characteristic combination with a common noise intensity interval; when there is a task characteristic combination, determining the highest priority combination in the task characteristic combination according to the task characteristic priority; and determining the common noise intensity interval of the highest priority combination as the target noise intensity interval.

[0010] In one example, the highest priority combination in the task characteristic combination is determined according to the task characteristic priority, specifically including: sorting the task characteristics in each task characteristic combination according to the order from high to low of the task characteristic priority; and performing a transitive comparison on the multiple task characteristic combinations to obtain the highest priority combination; the transitive comparison refers to comparing two task characteristic combinations, and then comparing the high priority task characteristic combination in the comparison result with the next other task characteristic combination again.

[0011] In one example, the transitive comparison is performed on the multiple task characteristic combinations to obtain the highest priority combination, specifically including: randomly selecting two task characteristic combinations from the multiple task characteristic combinations; starting from the first sorted position to perform a traversal comparison, and stopping the traversal comparison when the task characteristics at the same sorted position in the two task characteristic combinations are different or there is no same sorted position; when the task characteristics at the same sorted position are different, determining the task characteristic combination with the higher priority task characteristic as the high priority task characteristic combination; or when there is no same sorted position, determining the task characteristic combination with the remaining sorted position as the high priority task characteristic combination; comparing the high priority task characteristic combination with the next other task characteristic combination again until there is no next other task characteristic combination, and determining the high priority task characteristic combination in the last comparison result as the highest priority combination.

[0012] In one example, the method further includes: when there is no task characteristic combination, determining other task characteristics related to each overlapping interval; and determining the overlapping interval of the highest priority other task characteristic as the target noise interval.

[0013] In one example, the method further comprises: detecting the current model parameter of the global large model when the privacy consumption value exceeds the preset privacy budget; comparing the current model parameter with the original model parameter of the global large model to obtain a model parameter change value; and determining that the training task is in a privacy leakage state when the model parameter change value is greater than or equal to a preset parameter change value threshold.

[0014] In another aspect, the embodiments of the present application provide a noise data processing device based on federated learning, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any one of the above noise data processing methods based on federated learning.

[0015] In another aspect, the embodiments of the present application provide a noise data processing device based on federated learning, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any one of the above noise data processing methods based on federated learning.

[0016] The above at least one technical solution adopted by the embodiments of the present application can achieve the following beneficial effects:

[0017] The present application evaluates the strength of privacy protection through privacy budget, and dynamically adjusts the noise intensity according to the task characteristics, ensuring that the privacy of data in the training process is fully protected, avoiding the leakage of sensitive information, so that the global model can protect privacy while minimizing the negative impact of noise on the training effect, thereby improving the accuracy and generalization ability of the model. At the same time, the global model can adapt to diversified task requirements. Different training tasks have different requirements for privacy protection and model performance. Dynamic noise adjustment can flexibly respond to specific task characteristics and meet diversified needs. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the present application, some embodiments of the present application will be described in detail below in conjunction with the drawings, in which:

[0019] Figure 1 A flowchart of a noise data processing method based on federated learning provided by the embodiments of the present application is shown in the figure;

[0020] Figure 2 A noise intensity dynamic adjustment flowchart of a noise data processing method based on federated learning provided by the embodiments of the present application is shown in the figure;

[0021] Figure 3A privacy budget flowchart of a noise data processing method based on federated learning is provided for the embodiments of the present application.

[0022] Figure 4 A structural schematic diagram of a noise data processing device based on federated learning is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0023] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described below in detail with specific embodiments and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0024] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0025] Figure 1 A flowchart of a noise data processing method based on federated learning is provided for the embodiments of the present application. The method can be applied to different business fields. Some input parameters or intermediate results in the flow allow manual intervention adjustment to help improve accuracy.

[0026] The analysis method related to the embodiments of the present application can be implemented for a terminal device or a server, and the present application does not make special limitations thereon. For the convenience of understanding and description, the following embodiments are described in detail with a local server as an example.

[0027] Based on this, Figure 1 The flow in the method can include the following steps:

[0028] S101: Compare the privacy consumption value of the global large model with the preset privacy budget.

[0029] It should be noted that the differential privacy mechanism realizes privacy protection by adding noise to data or query results, and federated learning is to distribute the model training process to multiple local devices or data owners to avoid centralized transmission and storage of original data, thereby realizing privacy protection.

[0030] In the centralized processing scenario of the data set, differential privacy needs to add noise to the entire data set, etc., while in federated learning, differential privacy technology can be applied to each data subset on each local device, and then the model is aggregated, which can better adapt to the distributed data environment and reduce the impact of noise on model performance to some extent. For example: in a federated learning system composed of multiple mobile devices, each device stores user behavior data. Differential privacy processing can be performed on local data on each device before local model training, and then the training results are uploaded to the server for aggregation, thereby realizing privacy-protected model training in a distributed environment.

[0031] Privacy budget is a measure of the degree of data privacy protection. A smaller privacy budget value means that the algorithm protects privacy to a higher degree, as it limits the possibility of changes in the output results due to changes in individual data records. Intuitively, a smaller privacy budget value makes it more difficult for an attacker to infer information about a particular individual data by observing the output of the algorithm; while a larger privacy budget value indicates that greater privacy loss is allowed, i.e. the algorithm output is more likely to be affected by changes in individual data records, and the degree of privacy protection is relatively low.

[0032] In some embodiments of the present application, before comparing the privacy consumption value of the global large model with the preset privacy budget, task characteristics analysis is first required, including but not limited to tasks in the financial field, the technology field, the legal field, etc. The data distribution, feature type and privacy requirement of each task are different. The present application analyzes the task characteristics through the dimensions of data sensitivity, task complexity and training data volume, for example: some financial data (such as user's personal information, account history, etc.) is more sensitive and requires stronger privacy protection measures; a fraud detection task with high complexity may require more fine-grained noise adjustment; a task with a large amount of data is less sensitive to noise and can tolerate higher intensity noise to reduce the impact on performance.

[0033] Further, in order to realize dynamic noise adjustment, we integrate the differential privacy mechanism (Differential Privacy, DP) in the federated learning framework. Differential privacy adds Gaussian noise when updating the model. For high-sensitivity tasks, the noise intensity is large; for low-sensitivity tasks, the noise intensity is small. And according to the complexity and data volume of the task, different noise weighting strategies are used to prevent individual data leakage.

[0034] Further, in order to ensure the effectiveness of noise addition and the degree of privacy protection, a privacy evaluation module is introduced in the training process of federated learning to regularly evaluate the privacy protection level after adding noise. Through the privacy budget (privacy budget, ε) in the differential privacy theory, the privacy protection level of the model is evaluated, and the noise intensity is adjusted according to the evaluation result. ) to evaluate the strength of privacy protection, specifically: by tracking the privacy budget consumption of each model update, ensure that the preset privacy budget is not exceeded during training. Each model update should ensure that the privacy budget does not exceed the preset privacy budget = 1.

[0035] Evaluate the strength of privacy protection by privacy budget, ensure the privacy of data in the training process is fully protected, avoid sensitive information leakage.

[0036] S102: Stop updating the global large model when the privacy consumption value exceeds the preset privacy budget.

[0037] It should be noted that in some embodiments of the present application, if the privacy consumption value reaches the upper limit (for example = 1), the system will stop the model update of the current global large model, and when the privacy consumption value exceeds the preset privacy budget, detect the current model parameters of the global large model, evaluate whether too much personal data has been leaked, compare the current model parameters with the original model parameters of the global large model, get the model parameter change value, then set a parameter change value threshold (such as the model parameter change value is 0.1), when the model parameter change value is greater than or equal to the preset parameter change value threshold, it indicates that the privacy protection is insufficient, and it is determined that the training task is in a privacy leakage state.

[0038] Determine whether privacy is leaked by detecting the parameter change of the global large model, improve the privacy budget management and evaluation mechanism, and avoid the risk of privacy leakage.

[0039] S103: Find the task characteristics of the training task and the noise intensity interval of the task characteristics; each task characteristic has a corresponding noise intensity interval.

[0040] It should be noted that in some embodiments of the present application, for the task characteristics of the training task, specific noise intensity intervals are set, and different task characteristics correspond to different noise intensity intervals. For example:

[0041] For data sensitivity characteristics, high-sensitive data such as user personal information, account history, transaction records, etc., the noise intensity is set to 0.3 to 0.5; medium-sensitive data such as credit score, loan application record, etc., the noise intensity is set to 0.15 to 0.3; low-sensitive data such as publicly disclosed financial statement data, etc., the noise intensity is set to 0.05 to 0.1.

[0042] For the task complexity characteristic, high complexity tasks such as fraud detection involve complex models and a large number of features, and the noise intensity is set to 0.25-0.4; medium complexity tasks such as credit scoring involve relatively simple models, and the noise intensity is set to 0.15-0.25; low complexity tasks such as loan default prediction involve fewer features and simple models, and the noise intensity is set to 0.05-0.15.

[0043] For the data volume characteristic, large-scale data (>100,000 data), the noise intensity is set to 0.1-0.2; medium-scale data (10,000-100,000 data), the noise intensity is set to 0.2-0.3; small-scale data (<10,000 data), the noise intensity is set to 0.3-0.5.

[0044] By setting the noise intensity interval for the task characteristics, the dynamic noise adjustment can flexibly respond to the specific task characteristics and meet the diversified needs.

[0045] S104: Extracting a target noise intensity interval from the noise intensity interval according to the priority of the task characteristics of the training task.

[0046] It should be noted that before extracting the target noise intensity interval from the noise intensity interval, the priority of the task characteristics needs to be set, for example: data sensitivity > task complexity > data volume.

[0047] In some embodiments of the present application, when the training task has one task characteristic, the target noise intensity interval corresponding to the task characteristic of the training task is directly extracted. For example, the task characteristic of the training task is a high complexity task, and the target noise intensity interval extracted is 0.25-0.4.

[0048] When the training task has multiple task characteristics, find the overlapping interval between the noise intensity interval of the highest priority task characteristic and the noise intensity interval of other task characteristics to obtain an overlapping interval set; here, the other task characteristics are any one of the task characteristics possessed by the training task, i.e., each task characteristic possessed by the training task except the highest priority task characteristic.

[0049] When the overlapping interval set is empty, i.e., there is no overlapping area between the noise intensity interval of the highest priority task characteristic and the noise intensity interval of any other task characteristic, the noise intensity interval of the highest priority task characteristic is determined as the target noise interval. For example, the noise intensity interval of the highest priority task characteristic is (0, 1), and the noise intensity intervals of the other task characteristics are (1, 2) and (2, 3) respectively, and the target noise interval is (0, 1).

[0050] When the set of overlapping intervals is not empty, determine whether there is a task characteristic combination with common noise intensity intervals; here, the task characteristic combination with common noise intensity intervals is: if the noise intensity interval of the highest priority task characteristic of the training task has an overlapping interval with the noise intensity interval of any other task characteristic of the training task, then any task characteristic with an overlapping interval with the noise intensity interval of the highest priority task characteristic forms a task characteristic combination with the highest priority task characteristic; for example: the training task has three task characteristics A, B and C, A is the highest priority task characteristic, the noise intensity interval of A has an overlapping interval with the noise intensity interval of B, then (A, B) is a task characteristic combination; the noise intensity interval of A has a common overlapping interval with the noise intensity intervals of B and C, then (A, B, C) is a task characteristic combination; more specifically: the noise intensity interval of A is (0, 6), the noise intensity interval of B is (1, 5), the noise intensity interval of C is (2, 7), the common noise intensity interval of A, B and C is (2, 5), so A, B and C form a task characteristic combination (A, B, C).

[0051] Further, when there is a task characteristic combination, determine the highest priority combination in the task characteristic combination according to the task characteristic priority, and sort the task characteristics in each task characteristic combination according to the order of the task characteristic priority from high to low.

[0052] Perform a transitive comparison on the multiple task characteristic combinations to obtain the highest priority combination; the transitive comparison refers to comparing the results of two task characteristic combinations, and then comparing the high priority task characteristic combination in the comparison results with the next other task characteristic combination. The specific steps are: from the multiple task characteristic combinations, randomly select two task characteristic combinations, then compare the two task characteristic combinations from the first sorted position, stop the comparison when the task characteristics at the same sorted position in the two task characteristic combinations are different or do not have the same sorted position; when the task characteristics at the same sorted position are different, determine the task characteristic combination with the higher priority task characteristic as the high priority task characteristic combination; or when there is no same sorted position, determine the task characteristic combination with the remaining sorted position as the high priority task characteristic combination, then compare the high priority task characteristic combination with the next selected other task characteristic combination again, until there is no next selected other task characteristic combination, then determine the high priority task characteristic combination in the last comparison result as the highest priority combination.

[0053] For example, four task characteristics A, B, C and D, the priority is A > B > C > D, the task characteristic combination is (A, B, C), (A, C, D) and (A, B, C, D), any two task characteristic combinations (A, B, C) and (A, C, D) are selected, and the first sorting position is traversed and compared. Since the first sorting position task characteristics of the two characteristic combinations are A, continue to traverse, and the priority of B is higher than that of C, so the priority of the characteristic combination is (A, B, C) > (A, C, D), then (A, B, C) and (A, B, C, D) are compared from the first sorting position. Since the first three sorting positions of the two characteristic combinations are A, B and C, continue to traverse, and since (A, B, C) and (A, B, C, D) do not have the same sorting position in the next position, the task characteristic combination (A, B, C, D) with the remaining sorting position is determined as the highest priority task characteristic combination.

[0054] Finally, the common noise intensity interval of the highest priority combination is determined as the target noise intensity interval.

[0055] When there is no task characteristic combination, determine other task characteristics related to each overlapping interval, and determine the overlapping interval of the highest priority other task characteristic as the target noise interval. That is, only the highest priority task characteristic has an overlapping interval with a single other task characteristic, for example, three task characteristics A, B and C, and the overlapping interval is only A and B, A and C. Then only the task characteristic priority between B and C needs to be compared, and the priority B > C. Therefore, the target noise interval is the overlapping interval of A and B.

[0056] By setting the priority of the task characteristics and selecting the target noise interval, the global model can reduce the negative impact of noise on the training effect while protecting privacy, thereby improving the accuracy and generalization ability of the model.

[0057] S105: Adjust the added noise of the training task according to the target noise intensity interval.

[0058] It should be noted that in some embodiments of the present application, the noise added to the training task is adjusted within the target noise intensity interval, for example, the intelligent agent is trained using reinforcement learning. Set the state space, action space and reward function, initialize the deep reinforcement neural network DQN, collect experience through experience replay, train the network using the back propagation algorithm, so that the intelligent agent can adjust the noise intensity according to the task real-time state within the target noise intensity interval, and continuously learn optimization. After model training, during the training process, the model will continuously adjust its parameters through optimization algorithms (such as gradient descent, etc.) according to the characteristics and laws of local data. This parameter adjustment process is model updating. For example, in the credit scoring task, the client trains the model using local customer credit-related data, and the model's weight and bias parameters will change continuously to better fit the local data and improve the prediction accuracy of credit scoring.

[0059] After completing the model training, the model parameters updated by local training (usually with noise added to protect privacy) are uploaded to the server. The server collects model updates from various clients and processes them according to specific aggregation strategies (such as FedAvg aggregation strategy) to integrate them into the global model.

[0060] Through this method, the global model is optimized and improved, enabling it to integrate information and training results from various clients to improve overall performance and generalization ability.

[0061] It should be noted that although the embodiments of the present application are introduced and described in sequence with reference to Figure 1 steps S101 to S105, this does not mean that steps S101 to S105 must be executed in strict sequence. The reason why the embodiments of the present application introduce and describe steps S101 to S105 in sequence as shown in Figure 1 is to facilitate understanding of the technical solutions of the embodiments of the present application by those skilled in the art. In other words, in the embodiments of the present application, the sequence of steps S101 to S105 can be adjusted as needed.

[0062] Figure 1 In the method, the strength of privacy protection is evaluated by privacy budget, and the noise intensity is dynamically adjusted according to the characteristics of the task to ensure that the privacy of the data is fully protected during the training process, avoid sensitive information leakage, and make the global model reduce the negative impact of noise on the training effect as much as possible while protecting privacy, thereby improving the accuracy and generalization ability of the model. At the same time, the global model can adapt to diverse task requirements. Different training tasks have different requirements for privacy protection and model performance. Dynamic noise adjustment can flexibly respond to specific task characteristics to meet diverse needs.

[0063] Figure 2 A noise intensity dynamic adjustment flowchart of a noise data processing method based on federated learning provided by an embodiment of the present application.

[0064] In the method, first, task characteristics are analyzed, then a differential privacy mechanism is integrated into a federated learning framework, and then an evaluation result (whether privacy consumption exceeds a budget) is generated through an added evaluation module. Figure 2 If the privacy consumption of the evaluation result exceeds the privacy budget, the noise intensity is adjusted according to the task characteristics, and if the privacy consumption of the evaluation result does not exceed the privacy budget, the default noise intensity remains unchanged.

[0065] Figure 3 A privacy budget flowchart of a noise data processing method based on federated learning provided by an embodiment of the present application.

[0066] In the method, first, a privacy budget is initialized, then whether the privacy consumption of model updating exceeds the budget is observed, if the privacy consumption exceeds the budget, budget-exceeding processing (adjusting noise according to task characteristics) is performed, and privacy leakage detection is also performed, if privacy leakage is detected, privacy leakage processing (adjusting noise according to task characteristics) is performed, and then the privacy protection effect is evaluated. Figure 3 If no privacy leakage is detected, the privacy protection effect is directly evaluated, and if the privacy consumption does not exceed the budget, model updating is continued, and after the updating is completed, the privacy protection effect is evaluated.

[0067] Figure 4 A structural schematic diagram of a noise data processing device based on federated learning provided by an embodiment of the present application, comprising:

[0068] at least one processor; and

[0069] a memory in communication connection with the at least one processor; wherein

[0070] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any one of the noise data processing methods based on federated learning.

[0071] Some embodiments of the present application provide a non-volatile computer storage medium for noise data processing based on federated learning, which stores computer executable instructions, and the computer executable instructions can execute any one of the noise data processing methods based on federated learning.

[0072] The various embodiments in the present application are described in a progressive manner, and the same or similar parts among the various embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, the device and medium embodiments are described simply because they are substantially similar to the method embodiments.

[0073] The device and medium provided by the embodiments of the present application are one-to-one corresponding to the method, and therefore, the device and medium also have similar beneficial technical effects to the method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the device and medium will not be described here again.

[0074] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0075] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device implemented in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by the one or more blocks.

[0076] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implements the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0077] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0078] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0079] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the processor can execute instructions. The memory can also include non-volatile memory, such as read only memory (ROM), electrically programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), flash memory, or other memory technologies, CD-ROM, digital versatile disc (DVD), or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information for access by a computing device. In no case does the medium include a transitory signal.

[0080] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically programmable read only memory (EEPROM), flash memory or other memory technologies, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information that can be accessed by a computing device. According to the definition herein, computer readable media does not include transitory media such as modulated data signals and carrier waves.

[0081] It should also be noted that the terms "comprising," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article or apparatus that comprises the element.

[0082] ​​The above merely provides an example of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the technical principles of the present application should fall into the protection scope of the present application.

Claims

1. A noise data processing method based on federated learning, applied to a client-server distributed training architecture in a financial data sensitive field, characterized in that, The method comprises: During local model training of the federated learning client, detecting model parameter changes by a processor, comparing a privacy consumption value of the global large model with a preset privacy budget; When the privacy consumption value exceeds the preset privacy budget, stopping updating the global large model; Finding task characteristics of the training task and noise intensity intervals of the task characteristics; each task characteristic has a corresponding noise intensity interval; the data distribution, feature type and privacy requirement of different training tasks are different; the task characteristics include data sensitivity, task complexity and data volume; the higher the data sensitivity, the higher the noise intensity, the higher the task complexity, the higher the noise intensity, and the higher the data volume, the lower the noise intensity; According to the task characteristic priority of the training task, extracting a target noise intensity interval from the noise intensity intervals; According to the target noise intensity interval, adjusting the added noise of the training task to dynamically adjust the noise intensity added during local model training of the federated learning client; The method further comprises: When the privacy consumption value exceeds the preset privacy budget, detecting the current model parameters of the global large model; Comparing the current model parameters with the original model parameters of the global large model to obtain a model parameter change value; When the model parameter change value is greater than or equal to a preset parameter change value threshold, it is determined that the training task is in a privacy leakage state; Dynamically adjusting the noise intensity added during local model training of the federated learning client, specifically including: Based on a deep reinforcement learning algorithm, a noise intensity dynamic adjustment module is constructed, which optimizes the noise addition strategy in the target noise intensity interval in real time through training of an intelligent agent; By setting a state space, an action space and a reward function, and initializing a deep reinforcement neural network, experience is collected through experience replay, and the network is trained using a backpropagation algorithm, so that the intelligent agent can adjust the noise intensity according to the task real-time state in the target noise intensity interval, and continuously learn and optimize.

2. The method of claim 1, wherein, According to the task characteristic priority of the training task, extracting a target noise intensity interval from the noise intensity intervals, specifically including: When there are multiple task characteristics, find the overlapping interval between the noise intensity interval of the highest priority task characteristic and the noise intensity interval of other task characteristics to obtain an overlapping interval set; According to the overlapping interval set, extract a target noise intensity interval from the noise intensity interval of the highest priority task characteristic.

3. The method of claim 2, wherein, According to the overlapping interval set, extract a target noise intensity interval from the noise intensity interval of the highest priority task characteristic, specifically including: When the overlapping interval set is empty, the noise intensity interval of the highest priority task characteristic is determined as the target noise interval.

4. The method of claim 2, wherein, The method further comprises: When the overlapping interval set is not empty, determine whether there is a task characteristic combination with a common noise intensity interval; When there is a task characteristic combination, determine the highest priority combination in the task characteristic combination according to the task characteristic priority; The common noise intensity interval of the highest priority combination is determined as the target noise intensity interval.

5. The method of claim 4, wherein, The highest priority combination in the task characteristic combination is determined according to the task characteristic priority, and specifically comprises the following steps: The task characteristics in each task characteristic combination are sorted according to the order of the task characteristic priority from high to low; The multiple task characteristic combinations are compared in a transfer manner to obtain the highest priority combination; the transfer comparison refers to comparing the two task characteristic combinations, and then comparing the high priority task characteristic combination in the comparison result with the next other task characteristic combination.

6. The method of claim 5, wherein, The multiple task characteristic combinations are compared in a transfer manner to obtain the highest priority combination, and specifically comprises the following steps: Two task characteristic combinations are randomly selected from the multiple task characteristic combinations; The traversal comparison is started from the first sorting position, and when the task characteristics at the same sorting position in the two task characteristic combinations are different or do not have the same sorting position, the traversal comparison is stopped; When the task characteristics at the same sorting position are different, the task characteristic combination with the high task characteristic priority is determined as the high priority task characteristic combination; or when there is no same sorting position, the task characteristic combination with the remaining sorting position is determined as the high priority task characteristic combination; The high priority task characteristic combination is compared with the next other task characteristic combination, and the comparison is repeated until the next other task characteristic combination cannot be selected; the high priority task characteristic combination in the last comparison result is determined as the highest priority combination.

7. The method of claim 4, wherein, The method further comprises the following steps: When there is no task characteristic combination, other task characteristics related to each overlapping interval are determined; The overlapping interval of the highest priority other task characteristic is determined as the target noise interval.

8. A noisy data processing device based on federated learning, comprising: Comprise: At least one processor; And The memory is in communication connection with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the noise data processing method based on federated learning in any one of claims 1-7.

9. A non-volatile computer storage medium for processing noisy data based on federated learning, storing computer-executable instructions, characterized in that, The computer executable instructions can execute the noise data processing method based on federated learning in any one of claims 1-7.

Images