CN114358912A - An anomaly detection method based on federated learning and risk weight fusion - Google Patents

Abstract

The present invention provides an anomaly detection method based on federated learning and risk weight fusion. The method includes: treating each banking institution participating in federated learning as a client, each client establishes a machine learning model respectively, and in each round of iteration, after the machine learning model of each client is iteratively trained with a local data sample set, extracting parameters to update information and the risk weight information of this round of training, and upload them to the central server; the central server fuses all the received parameter update information with the risk weight information of each client for secure aggregation, and sends the joint model parameter update information to each client. , each client updates the parameters of the local machine learning model according to the received joint model update information. The invention strengthens the extraction of the advantageous features of the corresponding participants by means of risk weight aggregation, and improves the accuracy and recall rate of abnormal data detection in the field of abnormal detection such as financial fraud identification.

Description

An anomaly detection method based on federated learning and risk weight fusion

technical field

The invention relates to the technical field of financial data security detection, in particular to an anomaly detection method based on federated learning and risk weight fusion.

Background technique

With the popularity of the Internet, financial scenarios are more deeply integrated into people's daily lives to provide convenient services. However, financial fraud is also expanding using new technologies, causing huge losses to financial institutions and consumers. Traditional statistical and rule-based risk control methods cannot effectively detect changing fraud patterns. Machine learning and deep learning technologies have provided new ideas for financial transaction fraud detection, and their effectiveness has been verified in multi-domain scenarios. The detection effect of machine learning models usually relies on a large number of data sets. However, financial data is naturally sensitive and private, and data from different institutions cannot be directly collected and processed, which limits the application of machine learning technology in the financial field. Therefore, it is of great value to study how to train machine learning models under the condition of ensuring data privacy and security.

Financial data detection and application schemes based on privacy security include multi-party secure computing, homomorphic encryption and differential privacy, etc. These technologies ensure data privacy from the aspects of multi-party data security interaction, data encryption and data disturbance, but the data will still flow out of the local, there is leakage risk. Based on the fact that the data is kept locally in the organization, the model parameter aggregation method is used instead, which protects the privacy and security of user data in an innovative way. Some scholars further divide the extended application scope of federated learning into horizontal federated learning, vertical federated learning and federated transfer learning.

A method of applying federated learning to financial fraud detection in the prior art is: applying federated learning to credit card fraud detection, through joint multi-agency data training, compared with traditional fraud detection methods, the AUC index is improved by 10% . Applying longitudinal federated learning combined with a bounded constraint-based logistic regression algorithm to credit score prediction, the performance of AUC and Kolmogorov-Smirnov (KS) statistics is significantly improved due to the data enrichment brought by federated learning.

One of the above-mentioned methods of applying federated learning to financial fraud detection has the disadvantage that: this method treats each institutional model as an equal status based on the federated learning idea, uses the mean method to perform simple parameter aggregation, and does not target each institution. The individual characteristics of datasets and models are specifically analyzed and studied, which results in slow training and convergence of the central model, low sensitivity to fraud samples, and low fraud identification accuracy and recall.

SUMMARY OF THE INVENTION

Embodiments of the present invention provide an anomaly detection method based on federated learning and risk weight fusion, so as to effectively improve the effect of financial fraud detection.

In order to achieve the above objects, the present invention adopts the following technical solutions.

An anomaly detection method based on federated learning and risk weight fusion, including:

Each banking institution participating in federated learning is regarded as a client, each client establishes a machine learning model with the same structure, and the initial parameters of each machine learning model are issued by the central parameter server;

In each iteration, after the machine learning model of each client is iteratively trained using the local data sample set, the parameter update information and the risk weight information of this round of training are extracted, and the parameter update information and risk weight information are uploaded to the central server;

After the central server fuses all the received parameter update information with the risk weight information of each client for security aggregation, it issues the joint model parameter update information to each client, and each client performs a local machine according to the received joint model update information. The parameters of the learning model are updated, and each client uses its own machine learning model to detect anomalies in local data.

Preferably, each banking institution participating in the federated learning is regarded as a client, each client establishes a machine learning model with the same structure, and the initial parameters of each machine learning model are issued by the central parameter server, including:

Assume a fixed group of local banking institutions as participants in federated learning, consider each banking institution as a client, and each client establishes its own machine learning model. The structure of each machine learning model is the same. The initialization parameters sent are the same. During training, each client uses the central parameter server to securely aggregate and synchronize parameter information. Each client independently uses its own machine learning model to detect anomalies in local data. The machine learning models of all clients are iterated through communication. Form a central joint anomaly detection model.

Preferably, in each iteration, after the machine learning model of each client is iteratively trained using the local data sample set, the parameter update information and the risk weight information of this round of training are extracted, and the parameter update information and risk weight information are uploaded. to a central server, including:

Suppose fixed C banks participate in joint model training, each banking institution has a local data sample set, and the data sample set of the c-th banking institution is D _c :

是特征向量，

是标签，n_c表示参与联邦学习的第c个银行机构的数据集的大小，C为银行机构的总数；

is the eigenvector,

is the label, n _c represents the size of the dataset of the c-th banking institution participating in federated learning, and C is the total number of banking institutions;

计算本轮训练的风险权重信息：模型准确率

和针对欺诈严重程度的权重调节策略s_c，将

和s_c上传至中央参数服务器。In each round of training iteration t=1,2,..., each banking institution conducts machine learning model training based on its own data sample set, and calculates the parameter update information of this round of training

Calculate the risk weight information for this round of training: model accuracy

and a weight-adjusted strategy s _c for fraud severity, which will

and s _c are uploaded to the central parameter server.

计算本轮训练的风险权重信息：模型准确率

和针对欺诈严重程度的权重调节策略s_c，将

和s_c上传至中央参数服务器进行安全聚合，包括：Preferably, each banking institution conducts machine learning model training based on its own data sample set, and calculates the parameter update information for this round of training

Calculate the risk weight information for this round of training: model accuracy

and a weight-adjusted strategy s _c for fraud severity, which will

and s _c are uploaded to the central parameter server for secure aggregation, including:

代表第c家银行机构第t轮训练中获得的参数更新信息，具体计算公式如下：1):

Represents the parameter update information obtained in the t-th round of training of the c-th banking institution. The specific calculation formula is as follows:

代表本轮次第c家银行进行本地机器学习模型训练迭代生成的新的模型参数，各客户端端本地完成多次迭代，

代表完成整个联合训练轮次迭代的所有梯度更新之后的参数；Among them, w _t-1 represents the joint model parameters of the previous round,

On behalf of the c-th bank in this round, the new model parameters generated by the local machine learning model training iteration, each client completes multiple iterations locally,

Represents the parameters after completing all gradient updates for the entire joint training round iteration;

For each gradient descent training process, the parameters of the local machine learning model are updated gradiently according to the loss tested on the data sample set according to the learning rate η. The specific calculation formula is as follows:

代表第c家银行本地数据集训练时模型的平均损失对当前参数

的梯度，损失值L_c(x_c,y_c；w_t)由每个数据样本的损失在该银行数据集所有样本上求均值而得，具体计算公式如下：

Represents the average loss of the model when training on the local dataset of the cth bank versus the current parameters

The gradient of , the loss value L _c (x _c , y _c ; w _t ) is obtained by averaging the loss of each data sample over all samples in the bank data set, and the specific calculation formula is as follows:

Among them, l(x _i , _yi ; w _t ) represents the loss of a single sample, D _c represents the sample space of the data set of the c th bank, and n _c represents the number of data set samples of the c th bank;

2) Calculate the model accuracy

代表某家银行该轮次训练所得本地机器学习模型预测正确数量所占样本总量的比例，具体计算公式如下：Model accuracy

It represents the proportion of the correct number of local machine learning models predicted by a bank in this round of training to the total number of samples. The specific calculation formula is as follows:

代表模型正确预测出的欺诈样本数量，

代表模型错误预测为欺诈样本的数量，

代表模型正确预测出的良性样本数量，

代表模型错误预测为良性样本的数量；

represents the number of fraud samples correctly predicted by the model,

represents the number of samples that the model incorrectly predicted as fraudulent,

represents the number of benign samples correctly predicted by the model,

Represents the number of benign samples incorrectly predicted by the model;

Calculate the fraud severity value s _c of each client's local data set:

Each banking institution calculates the value of the severity of fraud samples in the local dataset as S(s ₁ , s ₂ ...... s _C ), and the fraud severity parameter s _C is positively related to the size of the dataset, the proportion of fraud samples contained in the dataset, and the amount of fraud;

Assuming that a total of C banking institutions participate in the federated training, when the central parameter server aggregates the parameter gradient information, the update of the local machine learning model of the c-th banking institution is substituted into the calculation method of the fraud degree weight value as follows:

代表所有参与联邦训练的银行数据集的欺诈级别总和，r_c为第c家银行数据集的欺诈严重程度权值；In formula (6), the numerator _sc represents the fraud severity level of the data set of the cth bank, and the denominator

Represents the sum of the fraud levels of all the bank datasets participating in the federation training, and rc is the fraud severity weight of the _c -th bank dataset;

3) After each round of training of each client, the central parameter server calculates the update information of the machine learning model of each client that fuses the risk weight information, namely:

代表第c家银行机构该轮次训练的风险权重，融合后面的本地机器学习模型参数更新信息

生成最终要聚合的欺诈检测模型参数更新信息

In formula (7),

Represents the risk weight of the c-th banking institution for this round of training, and integrates the following local machine learning model parameter update information

Generate the final aggregated fraud detection model parameter update information

Preferably, after the central server fuses all the received parameter update information with the risk weight information of each client for security aggregation, it sends the joint model parameter update information to each client, and each client updates the information according to the received joint model. Perform local machine learning model parameter updates, including:

In joint training, the learning objectives of the central joint anomaly detection model are:

Formula (8) represents minimizing the average loss of the joint model on all bank data sets, where l(x, y; w) represents the average loss of the joint model on all data sets, and n represents the total sample size of all bank data sets , l(x _i , y _i ; w) represents the loss value of the joint model on the i-th sample;

According to the distribution characteristics of the data set size in each local machine learning model, the optimization objective function of the central joint anomaly detection model is:

In formula (9), n _c represents the number of samples in the local dataset owned by the bank, n represents the sum of the number of samples in the dataset of all banks, and L _c (x _c , y _c ; w) represents the data of the joint model in the cth bank The average loss value over the samples of the set, defined as:

n _c represents the number of samples in the local dataset owned by the c-th bank, and l( _xi , y _i ; w) represents the loss value of the joint model on the i-th sample in the c-th bank’s dataset;

The central server fuses the received parameter update information uploaded by each client with the risk weight information for secure aggregation, and obtains the joint model parameter update information. The calculation formula is as follows:

代表本轮次训练第c家银行本地机器学习模型的风险权重，

代表第c家银行本轮次迭代产生的模型参数更新信息。In formula (11), w _t-1 is the joint model parameter of the previous round of training,

represents the risk weight of the local machine learning model of the c-th bank trained in this round,

Represents the model parameter update information generated by the current iteration of the cth bank.

w _t is the joint model parameter after this round of updating;

The central server sends the joint model parameter update information to each client, and each client updates the local machine learning model according to the received joint model parameter update information, and obtains the updated central joint anomaly detection model; the above process will perform T Iterates until the overall central joint anomaly detection model reaches the convergence index.

It can be seen from the technical solutions provided by the above embodiments of the present invention that the method of the present invention strengthens the extraction of the advantageous features of the corresponding participants by means of risk weight aggregation, and suppresses the extraction of the disadvantaged features of all parties, thereby improving the model training iteration process and process. Improve the performance of the final model. Especially in the field of anomaly detection such as financial fraud identification, the accuracy and recall rate of abnormal data detection are improved.

Additional aspects and advantages of the present invention will be set forth in part in the following description, which will become apparent from the following description, or may be learned by practice of the present invention.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

FIG. 1 is a schematic diagram of an implementation of an anomaly detection method based on federated learning risk weight fusion provided by an embodiment of the present invention;

FIG. 2 is a processing flowchart of an anomaly detection method based on federated learning risk weight fusion provided by an embodiment of the present invention.

Detailed ways

Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals refer to the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, but not to be construed as a limitation of the present invention.

It will be understood by those skilled in the art that the singular forms "a", "an", "the" and "the" as used herein can include the plural forms as well, unless expressly stated otherwise. It should be further understood that the word "comprising" used in the description of the present invention refers to the presence of stated features, integers, steps, operations, elements and/or components, but does not preclude the presence or addition of one or more other features, Integers, steps, operations, elements, components and/or groups thereof. It will be understood that when we refer to an element as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Furthermore, "connected" or "coupled" as used herein may include wirelessly connected or coupled. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as herein, are not to be taken in an idealized or overly formal sense. explain.

In order to facilitate the understanding of the embodiments of the present invention, the following will take several specific embodiments as examples for further explanation and description in conjunction with the accompanying drawings, and each embodiment does not constitute a limitation to the embodiments of the present invention.

The embodiment of the present invention is based on the risk weight aggregation algorithm architecture of Federated Learning (FL), solves the data island problem in the data training process, trains the business sensitive data of each institution locally, and uses the federated learning architecture to ensure data privacy. On the premise of building a central joint anomaly detection model, it realizes the fusion learning of multi-source sensitive data models. During the training process, the local machine learning model of each institution is built based on Convolutional Neural Networks (CNN), and a Risk Weight Federated Aggregation (called FedRWA) is proposed to aggregate the parameters of each local machine learning model, so that the central The joint anomaly detection model can accurately obtain the fraud risk characteristics of each local dataset and improve the model training effect.

A schematic diagram of the implementation of an anomaly detection method based on federated learning and risk weight fusion provided by an embodiment of the present invention is shown in FIG. 1 , and a specific processing flow is shown in FIG. 2 , including the following processing steps:

Step S10 : each banking institution participating in the federated learning is regarded as a client, each client establishes a machine learning model with the same network structure, and the central parameter server initializes the parameters of the machine learning model and distributes it to each client.

Step S20: In each round of iteration, after the machine learning model of each client uses the local data sample set to train for k times, the parameter update information is extracted, the parameter update information is fused with the local risk weight information, and encrypted and uploaded to the central server.

Step S30: After the central server securely aggregates all the received parameter update information and the risk weight information of each client, it sends the joint model update information to each client, and each client performs a local machine according to the received joint model update mechanism. Parameter updates for the learned model.

The iterative training process of the above steps S20 and S30 is repeatedly performed until the machine learning model of each client converges. Then, each client utilizes the trained machine learning model for local anomaly detection.

The above step S10 specifically includes: setting a fixed group of local banking institutions as participants in the federated learning, considering each banking institution as a client, each client establishing its own machine learning model, and the structure of each machine learning model is the same , the initialization parameters are the same, the difference is the specific parameters of the model during the training iteration process. Each client connects and communicates with the central parameter server through the communication network. During training, each client uses the central parameter server to safely aggregate and synchronize parameter information.

Each client independently uses its own machine learning model for anomaly detection of local data, and the machine learning models of all clients will form a central joint anomaly detection model through communication iterations in subsequent steps.

The above-mentioned step S20 specifically includes: setting fixed C banks to participate in the joint model training, each banking institution has a local data sample set, and the data sample set of the c-th banking institution is D _c :

是特征向量，

是标签，n_c表示参与联邦学习的第c个银行机构的数据集的大小，C为银行机构的总数。

is the eigenvector,

is the label, n _c represents the size of the dataset of the c-th banking institution participating in federated learning, and C is the total number of banking institutions.

Since the datasets held by different banking institutions tend to have different characteristics, i.e. they have different dataset sizes, different numbers of fraud labels, different fraud severity, etc. Therefore, a federated training model parameter aggregation strategy based on risk weight adjustment is introduced to better capture the characteristics of fraud samples from different banking institutions. Mainly include the following:

的权重调节策略：1) For model accuracy

The weight adjustment strategy of:

作为权值，有效提取训练良好的本地机器学习模型信息，并加速整体模型的迭代收敛。When the joint model aggregates parameter information, the model accuracy obtained by each local machine learning model in this round of training based on local data testing

As weights, it effectively extracts well-trained local machine learning model information and accelerates the iterative convergence of the overall model.

2) Weight adjustment strategy _rc for fraud severity:

Due to the different number, frequency and amount of fraudulent attacks suffered by various banking institutions in their daily business, the collected data sets are distributed with different numbers and degrees of fraudulent user samples. Banks can still deal with small and simple fraud risks, but have zero tolerance for fraudulent behaviors such as gang large-value fraud that have serious impacts. Therefore, we propose a risk weight adjustment strategy _rc based on the severity of fraud, so that the joint model will focus on absorbing the model generated by the training of serious fraud data samples during the training process, so as to create a shared model that can identify serious fraud, and let all participants in the federation training Banking institutions benefit from it.

The severity of fraud in the dataset samples is divided into risk levels or weight rules, which can be set according to the actual business scenarios of financial institutions. The following is a classification method:

The risk level of the data sample set of each banking institution is defined as five levels:

Level 1: There are no fraudulent samples in the data set, or there are only a small amount of fraudulent samples, such as short-term overdue samples;

Level 2: There are a small number of fraudulent samples in the data set, and the amount of fraud is small, which is a daily credit risk that banking operations can deal with.

Level 3: There are many fraud samples in the data set, and the amount of fraud is small, but the bank needs to pay attention to and prevent fraud risks.

Level 4: There are many fraud samples in the data set, and there are individual cases with a large amount of fraud, which have a certain impact on the banking business and need to focus on analysis and prevention.

Level 5: The data set contains gang fraud and other sample cases with a large number, a large amount, and frequent occurrences, which have a serious impact on the corresponding business, and need to be traced to the source of the problem and investigated and prevented.

并计算本轮训练的风险权重信息包括本地模型准确率

和数据集欺诈严重程度s_c，融合

s_c作为欺诈检测模型参数更新信息上传至中央参数服务器。The parameter update information t of each banking institution's fraud detection model is determined according to the parameter update information and its risk weight obtained by training the local machine learning model on the data sample set of each banking institution. In each round of communication iteration t=1,2,..., each banking institution conducts machine learning model training based on its own data sample set, and calculates the parameter update information of this round of training

And calculate the risk weight information of this round of training, including the local model accuracy

and dataset fraud severity _sc , fused

s _c is uploaded to the central parameter server as the parameter update information of the fraud detection model.

代表第c家银行机构第t轮训练中获得的参数更新信息，如公式所示。1):

Represents the parameter update information obtained in the t-th round of training for the c-th banking institution, as shown in the formula.

代表本轮次第c家银行进行本地机器学习模型训练迭代生成的新的模型参数。根据本地机器学习模型训练时随机梯度下降批次大小B和数据集迭代次数E的不同，

代表完成一整个训练轮次迭代的所有梯度更新之后的参数。Among them, w _t-1 represents the joint model parameters of the previous round,

New model parameters generated iteratively by local machine learning model training on behalf of the c-th bank in this round. According to the difference between the stochastic gradient descent batch size B and the number of dataset iterations E when the local machine learning model is trained,

Represents the parameters after completing all gradient updates for an entire training epoch.

For each gradient descent training process, the local machine learning model parameters are gradient updated at the learning rate η based on the loss tested on the private dataset. As the formula shows:

代表第c家银行本地数据集训练时模型的平均损失对当前参数

的梯度，损失值L_c(x_c,y_c；w_t)由每个数据样本的损失在该银行数据集所有样本上求均值而得，如公式所示。

Represents the average loss of the model when training on the local dataset of the cth bank versus the current parameters

The gradient of , the loss value L _c (x _c , y _c ; w _t ) is obtained by averaging the loss of each data sample over all samples in the bank dataset, as shown in the formula.

Among them, l(x _i , _yi ; w _t ) represents the loss of a single sample, D _c represents the dataset sample space of the cth bank, and _nc represents the number of dataset samples of the cth bank.

2) The risk weight information for each round of training for each client is described as follows:

a) Calculate the model accuracy

代表某家银行该轮次训练所得本地机器学习模型预测正确数量所占样本总量的比例。具体计算公式如下：Model accuracy

It represents the proportion of the correct number of predictions made by the local machine learning model of a bank in this round of training to the total sample. The specific calculation formula is as follows:

代表模型正确预测出的欺诈样本数量，

代表模型错误预测为欺诈样本的数量，

代表模型正确预测出的良性样本数量，

代表模型错误预测为良性样本的数量。

represents the number of fraud samples correctly predicted by the model,

represents the number of samples that the model incorrectly predicted as fraudulent,

represents the number of benign samples correctly predicted by the model,

Represents the number of samples that the model incorrectly predicted as benign.

_b ) Calculate the fraud (or anomaly) severity level s _c and weight rc of each client's local dataset:

Suppose the severity of the fraud samples in the dataset of each banking institution is s(s ₁ , s ₂ ...... s _C ), and a total of C banking institutions participate in the federal training, then when the joint model aggregates the parameter gradient information, then the c th banking institution The update of the local machine learning model is substituted into the calculation method of the fraud degree weight value as follows:

代表所有参与联邦训练的银行数据集的欺诈级别总和。r_c即为第c家银行数据集的欺诈严重程度权值。In formula (6), the numerator _sc represents the fraud severity level of the data set of the cth bank, and the denominator

Represents the sum of fraud levels across all datasets of banks participating in federation training. _rc is the fraud severity weight of the cth bank dataset.

For example, 3 banks participate in federated training, and their local datasets are marked with serious fraud procedures as level 1, level 3 and level 5, then the aggregated local machine learning model of the second bank will substitute the weight value

r ₂ =3/1+3+5=1/3.

3) After each round of training of each client, the central parameter server calculates and aggregates the local machine learning model update information fused with risk weight information. which is:

代表第c家银行机构该轮次训练的风险权重，融合后面的本地机器学习模型参数更新信息

生成最终联合模型进行聚合的欺诈检测模型参数更新信息

In formula (7),

Represents the risk weight of the c-th banking institution for this round of training, and integrates the following local machine learning model parameter update information

Fraud detection model parameter update information for generating the final joint model for aggregation

The above step S30 specifically includes: in the joint training, the learning objectives of the central joint anomaly detection model are:

Formula (8) represents minimizing the average loss of the joint model on all bank data sets, where l(x, y; w) represents the average loss of the joint model on all data sets, and n represents the total sample size of all bank data sets , l(x _i , y _i ; w) represents the loss value of the joint model on the i-th sample.

According to the distribution characteristics of the data set size in each local machine learning model, the optimization objective function of the central joint anomaly detection model can be rewritten as:

In formula (9), n _c represents the number of local data set samples owned by the bank, and n represents the sum of the data set samples of all banks. L _c (x _c , y _c ; w) represents the average loss value of the joint model over the samples of the c-th bank dataset. defined as:

n _c represents the number of samples in the local dataset owned by the c-th bank, and l( _xi , y _i ; w) represents the loss value of the joint model on the i-th sample in the c-th bank's dataset.

In this method, the central server fuses the received parameter update information uploaded by each client with the risk weight information for security aggregation, obtains the joint model parameter update information, and issues the joint model parameter update information to each client. The local machine learning model is updated according to the received joint model parameter update information, and an updated central joint anomaly detection model is obtained.

代表本轮次训练第c家银行本地机器学习模型的风险权重，

代表第c家银行本轮次迭代产生的模型参数更新信息。w_t即为本轮次更新后的联合模型参数。In formula (11), w _t-1 is the joint model parameter of the previous round of training,

represents the risk weight of the local machine learning model of the c-th bank trained in this round,

Represents the model parameter update information generated by the current iteration of the cth bank. w _t is the joint model parameter after this round of updating.

The above process will go through T rounds of iterations until the overall central joint anomaly detection model reaches the convergence index.

The above algorithm is recorded as: Risk Weight Federated Aggregation (Risk Weight Federated Aggregation, called FedRWA).

To sum up, compared with the prior art that treats all parties equally during joint model training, the method of the present invention can specifically analyze the characteristics of each participant's model and data set during the joint model training process. The aggregation of risk weights strengthens the extraction of the advantageous features of the corresponding participants and suppresses the extraction of the disadvantaged features of all parties, thereby improving the model training iteration process and improving the performance of the final model. Especially in the field of anomaly detection such as financial fraud identification, the accuracy and recall rate of abnormal data detection are improved.

来进行多模型参数的加权聚合。这样在模型参数聚合时能有效提取各机构中训练效果较好的模型参数更新信息，加速整体模型训练的迭代收敛，缩短训练所需时间。The embodiment of the present invention is aimed at the specific performance of each institutional model in the training process, such as the local data verification accuracy rate

to perform weighted aggregation of multiple model parameters. In this way, when the model parameters are aggregated, the update information of the model parameters with better training effect in each institution can be effectively extracted, the iterative convergence of the overall model training can be accelerated, and the time required for training can be shortened.

The embodiment of the present invention analyzes the statistical characteristics and fraud severity characteristics of fraud samples in each institution's data set, defines a fraud risk weight classification model, and performs biased aggregation of each institution model parameter according to different risk weights. By accurately extracting the characteristics of fraud samples from each institution's datasets for model parameter aggregation, the ability of the central joint anomaly detection model to identify fraud samples can be effectively improved, thereby improving the effect of financial fraud detection and effectively detecting large-amount fraud samples.

Those of ordinary skill in the art can understand that the accompanying drawing is only a schematic diagram of an embodiment, and the modules or processes in the accompanying drawing are not necessarily necessary to implement the present invention.

From the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solutions of the present invention can be embodied in the form of software products in essence or the parts that make contributions to the prior art. The computer software products can be stored in storage media, such as ROM/RAM, magnetic disks, etc. , CD, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the apparatus or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and reference may be made to some descriptions of the method embodiments for related parts. The apparatus and system embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, It can be located in one place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. Substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (5)

1. An anomaly detection method based on risk weight fusion of federal learning is characterized by comprising the following steps:

taking each bank organization participating in federal learning as a client, respectively establishing machine learning models with the same structure by each client, and issuing initial parameters of each machine learning model by a central parameter server;

in each iteration, after the machine learning model of each client side is iteratively trained by using a local data sample set, extracting parameter updating information and risk weight information of the current iteration of training, and uploading the parameter updating information and the risk weight information to a central server;

and the central server fuses all the received parameter updating information with the risk weight information of each client for safety aggregation, and then sends the combined model parameter updating information to each client, each client updates the parameters of the local machine learning model according to the received combined model updating information, and each client performs anomaly detection on local data by using the own machine learning model.

2. The method according to claim 1, wherein the step of regarding each banking institution participating in federal learning as a client, each client respectively establishes a machine learning model with the same structure, and initial parameters of each machine learning model are issued by a central parameter server, comprises:

a group of local bank organizations are set as participants of federal learning, the bank organizations are regarded as clients, the clients respectively establish own machine learning models, the structures of the machine learning models are the same, initialization parameters issued by a central parameter server are the same, the clients perform parameter information safety aggregation and synchronization by means of the central parameter server during training, each client independently utilizes the own machine learning model to perform local data anomaly detection, and the machine learning models of all the clients form a central combined anomaly detection model through communication iteration.

3. The method of claim 1, wherein in each iteration, after the machine learning model of each client is iteratively trained by using the local data sample set, extracting parameter update information and risk weight information of the current round of training, and uploading the parameter update information and the risk weight information to the central server, the method comprises:

c fixed banks are set to participate in the joint model training, each bank organization has a local data sample set, and the data sample set of the C bank organization is D_c：

Is a vector of the features of the image,

is a label, n_cRepresenting the size of a data set of the C-th banking institution participating in federal learning, wherein C is the total number of the banking institutions;

in each round of training iteration t being 1,2, …, each banking institution performs machine learning model training on the basis of own data sample set, and calculates parameter updating information of the round of training

Calculating the risk weight information of the training of the current round: model accuracy

And a weight adjustment policy s for the severity of fraud_cWill be

And s_cAnd uploading to a central parameter server.

4. The method of claim 3, wherein the step of removing the substrate comprises removing the substrate from the substrateEach bank organization conducts machine learning model training on the basis of own data sample set and calculates the parameter updating information of the current training

Calculating the risk weight information of the training of the current round: model accuracy

And a weight adjustment policy s for the severity of fraud_cWill be

And s_cUploading to a central parameter server for security aggregation, comprising:

1)：

representing the parameter updating information obtained in the t round of training of the c-th banking institution, the specific calculation formula is as follows:

wherein, w_t-1Representing the parameters of the combined model in the previous round,

representing the new model parameters generated by local machine learning model training iteration performed by the c-th bank in the turn, each client locally completes multiple iterations,

representing the parameters after all gradient updates for the entire joint training round iteration are completed;

for each gradient descent training process, the local machine learning model parameters are subjected to gradient updating according to the loss tested on the data sample set and the learning rate eta, and the specific calculation formula is as follows:

representing the average loss of the model to the current parameter when the c-th bank local data set is trained

Gradient of (2), loss value L_c(x_c,y_c；w_t) The loss of each data sample is averaged over all samples of the bank data set, and the specific calculation formula is as follows:

wherein, l (x)_i,y_i；w_t) Representing the loss of a single sample, D_cSample space of data set, n, representing the c-th bank_cA data set sample number representing a c-th bank;

2) calculating model accuracy

Model accuracy

Representing the proportion of the predicted correct number of the local machine learning model obtained by the training of the round of a certain bank to the total amount of the samples, the specific calculation formula is as follows:

representing the number of fraud samples correctly predicted by the model,

the representative model mispredicts to the number of rogue samples,

representing the number of benign samples that the model correctly predicts,

representing the number of samples the model mispredicted as benign;

calculating fraud severity value s of each client local data set_c：

Each banking institution calculates a local data set fraud sample severity value S (S)₁,s₂……s_C) Fraud severity parameter s_CPositively correlated with the size of the data set, the proportion of fraud samples contained in the data set and the amount of fraud;

if a C bank institution participates in federal training, when a central parameter server aggregates parameter gradient information, the calculation method of substituting the update of the local machine learning model of the C bank institution into the fraud degree weight value is as follows:

in formula (6), the molecule s_cRepresenting the fraud severity level, denominator, of the c-th bank data set

Sum of fraud levels, r, representing all bank data sets participating in federal training_cThe fraud severity weight of the data set of the c-th bank;

3) after each round of training of each client, the central parameter server calculates the machine learning model update information of each client fusing the risk weight information, namely:

in the formula (7), the first and second groups,

representing the risk weight of the c-th banking institution in the round of training, fusing the parameter updating information of the subsequent local machine learning model

Generating fraud detection model parameter updates to be aggregated eventually

5. The method according to claim 3 or 4, wherein the central server performs security aggregation by fusing all the received parameter update information with risk weight information of each client, and then issues the associated model parameter update information to each client, and each client performs parameter update of the local machine learning model according to the received associated model update information, including:

in the joint training, the learning targets of the central joint anomaly detection model are as follows:

equation (8) represents minimizing the average loss of the federated model over all the bank data sets, where l (x, y; w) represents the average loss of the federated model over all the data sets, n represents the number of samples aggregated over all the bank data sets, and l (x)_i,y_i(ii) a w) represents the loss value of the combined model above the ith sample;

according to the distribution characteristics of the data set size in each local machine learning model, the optimization objective function of the central combined anomaly detection model is as follows:

n in formula (9)_cRepresenting the number of local data set samples owned by the bank, n representing the sum of the number of data set samples for all banks, L_c(x_c,y_c(ii) a w) represents the average loss value of the federation model over the sample of the set of family c bank data, defined as:

n_crepresenting the number of local data set samples owned by the c-th bank, l (x)_i,y_i(ii) a w) represents the loss value of the federation model on the ith sample in the c bank dataset;

the central server carries out safety aggregation on the received parameter updating information uploaded by each client and the risk weight information to obtain the parameter updating information of the combined model, and the calculation formula is as follows:

w in formula (11)_t-1Is the combined model parameter of the previous round of training,

representing the risk weight of the local machine learning model of the c-th bank in the training round,

and representing the updated information of the model parameters generated by the current iteration of the c-th bank. w is a_tThe parameters are the parameters of the combined model after the updating of the round;

the central server sends the combined model parameter updating information to each client, and each client updates a local machine learning model according to the received combined model parameter updating information to obtain an updated central combined anomaly detection model; the above process will perform T-round iteration until the overall central combined anomaly detection model reaches the convergence index.

Images