WO2024071728A1

WO2024071728A1 - Anomaly detection device using neural network learned using heterogeneous loss function

Info

Publication number: WO2024071728A1
Application number: PCT/KR2023/013473
Authority: WO
Inventors: 유광선; 이준기
Original assignee: 주식회사 엘로이랩
Priority date: 2022-09-30
Filing date: 2023-09-08
Publication date: 2024-04-04
Also published as: KR102506991B1; KR102506988B1

Abstract

An anomaly detection device is disclosed. The anomaly detection device using a heterogeneous loss function, according to one embodiment, comprises: a receiver for receiving data from which normality or anomaly is to be detected; and a processor for detecting whether an anomaly exists in the data, by inputting the data into a pre-learned neural network, wherein the pre-learned neural network can be learned through a heterogeneous loss function calculated using a gradient parameter determined on the basis of an anomaly score.

Description

Anomaly detection device using a neural network learned using a heterogeneous loss function

The following embodiments relate to an anomaly detection device using a neural network learned using a heterogeneous loss function.

Anomaly detection (or, anomaly detection) means identifying outliers or abnormalities in data. Anomaly detection assumes that a model (e.g., a neural network model) is trained on a dataset consisting of only normal data, so that the output contains only normal features.

Autoencoder is a representative model where the output is generated by an encoder and decoder. The encoder converts the input into a latent vector, and the latent vector is restored to the original by the decoder.

Due to a normal dataset, the output of the autoencoder has normal features, so the reconstruction error (difference between input and output) is close to 0 for normal inputs, while it is close to 0 for abnormal (or abnormal) inputs. It has a high value.

However, there is a problem that it is difficult to create a clean dataset due to the ambiguity between normal and abnormal data. Even though the data contains both normal and abnormal data, contaminated data may be generated when the training data is labeled as normal. Because contaminated data can degrade model performance, there is a need to build a model that is robust to contamination.

Embodiments may provide an anomaly detection technology using a heterogeneous loss function.

However, technical challenges are not limited to the above-mentioned technical challenges, and other technical challenges may exist.

An anomaly detection device according to an embodiment includes a receiver that receives data to detect whether it is normal or abnormal; and a processor that detects anomalies in the data by inputting the data into a pre-trained neural network, wherein the pre-trained neural network has a gradient parameter determined based on the anomaly score of the training data. It can be learned through a heterogeneous loss function calculated using ).

The abnormality score according to one embodiment may be calculated based on the reconstruction error.

The gradient parameter may be calculated based on a z-score calculated based on the median of the learning data and the maximum absolute value of the z-score.

The z-score may be calculated based on the Median Absolute Deviation (MAD) of the learning data, which is calculated based on the median of the learning data.

The gradient parameter may be determined based on the larger value of a predetermined constant and the maximum value.

The heterogeneous loss function is determined based on Equation 1 to Equation 3, and Equation 1 is:

And Equation 2 is,

And Equation 3 above is,

ego,

represents the training data,

is the average of the training data, N is the number of training data, c is the first hyperparameter,

represents the ideal score,

is the z score,

may refer to the parameter for the lowest gradient loss.

A method of anomaly detection according to an embodiment includes receiving data to detect whether normal or abnormality is detected; And detecting whether the data is abnormal by inputting the data into a pre-trained neural network, wherein the pre-trained neural network has a gradient parameter determined based on the abnormality score of the training data. It can be learned based on the heterogeneous loss function calculated through .

A learning method using a heterogeneous loss function according to an embodiment includes receiving learning data; and determining a gradient parameter based on the abnormality score of the learning data; determining a heterogeneous loss function based on the gradient parameters; And it may include training an artificial neural network based on the heterogeneous loss function.

1 shows a schematic block diagram of an abnormality detection device according to an embodiment.

Figure 2 shows an example of the distribution of anomaly scores of learning data.

Figure 3 shows examples of data corruption rates according to different displacements.

Figure 4 shows an example of a learning algorithm for a neural network used in the abnormality detection device of Figure 1.

Figure 5 shows an example of a graph for explaining a heterogeneous loss function.

Figure 6 shows another example of a graph for explaining a heterogeneous loss function.

Figure 7 is a table showing AUROC (Area Under the Receiver Operating Characteristic) according to various loss functions.

8 to 10 show an example of AUROC (Area Under the Receiver Operating Characteristic) for the contamination rate of a plurality of models.

11 to 13 show other examples of AUROC for pollution rates of multiple models.

14 to 16 show another example of AUROC for contamination rates of multiple models.

Figure 17 is a table showing AUROC according to various neural network models.

Figures 18 to 20 show examples of comparing the AUROC for the contamination rate of the neural network model of the plurality of models in Figure 1.

Figure 21 shows the results of an elimination study for heterogeneous loss functions;

Figure 22 is a graph to explain the sensitivity of heterogeneous loss functions and soft rejection.

FIG. 23 shows a flowchart of the operation of the abnormality detection device shown in FIG. 1.

FIG. 24 shows a flowchart of an operation for training the anomaly detection device shown in FIG. 1.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in this specification are merely illustrative for the purpose of explaining the embodiments according to the concept of the present invention. They may be implemented in various forms and are not limited to the embodiments described herein.

Since the embodiments according to the concept of the present invention can make various changes and have various forms, the embodiments will be illustrated in the drawings and described in detail in this specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes changes, equivalents, or substitutes included in the spirit and technical scope of the present invention.

Terms such as first or second may be used to describe various components, but the components should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another component, for example, a first component may be named a second component, without departing from the scope of rights according to the concept of the present invention, Similarly, the second component may also be referred to as the first component.

When a component is said to be "connected" or "connected" to another component, it is understood that it may be directly connected to or connected to the other component, but that other components may exist in between. It should be. On the other hand, when it is mentioned that a component is “directly connected” or “directly connected” to another component, it should be understood that there are no other components in between. Expressions that describe the relationship between components, such as “between”, “immediately between” or “directly adjacent to”, should be interpreted similarly.

The terminology used herein is only used to describe specific embodiments and is not intended to limit the invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as "include" or "have" are intended to designate the presence of a described feature, number, step, operation, component, part, or combination thereof, and one or more other features or numbers, It should be understood that this does not exclude in advance the possibility of the presence or addition of steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the technical field to which the present invention pertains. Terms as defined in commonly used dictionaries should be interpreted as having meanings consistent with the meanings they have in the context of the related technology, and unless clearly defined in this specification, should not be interpreted in an idealized or overly formal sense. No.

A module in this specification may mean hardware that can perform functions and operations according to each name described in this specification, or it may mean computer program code that can perform specific functions and operations. , or it may mean an electronic recording medium loaded with computer program code that can perform specific functions and operations, for example, a processor or microprocessor.

In other words, a module may mean a functional and/or structural combination of hardware for carrying out the technical idea of the present invention and/or software for driving the hardware.

Hereinafter, embodiments will be described in detail with reference to the attached drawings. However, the scope of the patent application is not limited or limited by these examples. The same reference numerals in each drawing indicate the same members.

Referring to FIG. 1, the anomaly detection device 10 can detect anomalies in data. The abnormality detection device 10 can detect whether data is normal and/or abnormal (or abnormal).

Data may consist of information in a form that can be processed by a computer. Data can be in the form of letters, numbers, sounds, pictures, etc. that a computer can process. For example, data may include images. An image includes an image of an object created by refraction or reflection of light, and may mean representing the shape of the object using lines or colors.

The anomaly detection device 10 can detect anomalies in data using a neural network. The anomaly detection device 10 can detect anomalies in data by training a neural network based on learning data and processing the data based on the learned neural network.

Neural networks (or artificial neural networks) can include statistical learning algorithms that mimic neurons in biology in machine learning and cognitive science. A neural network can refer to an overall model in which artificial neurons (nodes), which form a network through the combination of synapses, change the strength of the synapse connection through learning and have problem-solving capabilities.

Neurons in a neural network can contain combinations of weights or biases. A neural network may include one or more layers consisting of one or more neurons or nodes. Neural networks can infer the results they want to predict from arbitrary inputs by changing the weights of neurons through learning.

Neural networks may include deep neural networks. Neural networks include CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), perceptron, multilayer perceptron, FF (Feed Forward), RBF (Radial Basis Network), DFF (Deep Feed Forward), and LSTM. (Long Short Term Memory), GRU (Gated Recurrent Unit), AE (Auto Encoder), VAE (Variational Auto Encoder), DAE (Denoising Auto Encoder), SAE (Sparse Auto Encoder), MC (Markov Chain), HN (Hopfield) Network), BM (Boltzmann Machine), RBM (Restricted Boltzmann Machine), DBN (Depp Belief Network), DCN (Deep Convolutional Network), DN (Deconvolutional Network), DCIGN (Deep Convolutional Inverse Graphics Network), GAN (Generative Adversarial Network) ), Liquid State Machine (LSM), Extreme Learning Machine (ELM), Echo State Network (ESN), Deep Residual Network (DRN), Differential Neural Computer (DNC), Neural Turning Machine (NTM), Capsule Network (CN), It may include Kohonen Network (KN) and Attention Network (AN).

The abnormality detection device 10 may be implemented with a printed circuit board (PCB) such as a motherboard, an integrated circuit (IC), or a system on chip (SoC). For example, the abnormality detection device 10 may be implemented as an application processor.

Additionally, the abnormality detection device 10 may be implemented in a personal computer (PC), a data server, or a portable device.

Portable devices include laptop computers, mobile phones, smart phones, tablet PCs, mobile internet devices (MIDs), personal digital assistants (PDAs), and enterprise digital assistants (EDAs). , digital still camera, digital video camera, portable multimedia player (PMP), personal navigation device or portable navigation device (PND), handheld game console, e-book ( It can be implemented as an e-book) or a smart device. A smart device may be implemented as a smart watch, smart band, or smart ring.

The abnormality detection device 10 includes a receiver 100 and a processor 200. The abnormality detection device 10 may further include a memory 300.

The receiver 100 can receive data to detect whether it is normal or abnormal. The receiver 100 may receive data from the outside or the memory 300. Receiver 100 may include a receiving interface. The receiver 100 may output the received data to the processor 200.

The processor 200 may process data stored in the memory 300. The processor 200 may execute computer-readable code (eg, software) stored in the memory 300 and instructions triggered by the processor 200 .

The “processor 200” may be a data processing device implemented in hardware that has a circuit with a physical structure for executing desired operations. For example, the intended operations may include code or instructions included in the program.

For example, data processing devices implemented in hardware include microprocessors, central processing units, processor cores, multi-core processors, and multiprocessors. , ASIC (Application-Specific Integrated Circuit), and FPGA (Field Programmable Gate Array).

The processor 200 may perform anomaly detection based on a neural network learned using a heterogeneous loss function.

The processor 200 may directly train a neural network used to detect outliers, but may also detect outliers using a separately learned neural network. In the process of training the neural network, the processor 200 may determine a gradient parameter based on the abnormality score of the training data received from the receiver 100. The processor 200 may calculate an anomaly score based on the restoration error for the learning data.

The processor 200 may calculate the z-score based on the median of the training data. The processor 200 may calculate the Median Absolute Deviation (MAD) of the data based on the median of the training data. Processor 200 may calculate the z-score based on the MAD.

The processor 200 may determine the gradient parameter based on the z score and the maximum absolute value of the z score. The processor 200 may determine the gradient parameter based on the larger value between a predetermined constant and the maximum absolute value of the z-score.

The processor 200 may determine a heterogeneous loss function based on the gradient parameter.

When the gradient parameter is the first value, the processor 200 may determine the first loss function to be a heterogeneous loss function. When the gradient parameter is the second value, the processor 200 may determine the second loss function to be a heterogeneous loss function.

The processor 200 may determine a first loss function and a second loss function based on a value divided by the data by the first hyperparameter.

The processor 200 can detect whether the input data is normal or abnormal by inputting data into a neural network learned based on a heterogeneous loss function.

The memory 300 may store instructions (or programs) executable by the processor 200. For example, the instructions may include instructions for executing the operation of the processor and/or the operation of each component of the processor.

The memory 300 may be implemented as a volatile memory device or a non-volatile memory device.

Volatile memory devices may be implemented as dynamic random access memory (DRAM), static random access memory (SRAM), thyristor RAM (T-RAM), zero capacitor RAM (Z-RAM), or twin transistor RAM (TTRAM).

Non-volatile memory devices include EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, MRAM (Magnetic RAM), Spin-Transfer Torque (STT)-MRAM (MRAM), and Conductive Bridging RAM (CBRAM). , FeRAM (Ferroelectric RAM), PRAM (Phase change RAM), Resistive RAM (RRAM), Nanotube RRAM (Nanotube RRAM), Polymer RAM (PoRAM), Nano Floating Gate Memory (NFGM), holographic memory, molecular electronic memory device, or insulation resistance change memory.

Figure 2 shows an example of the distribution of anomaly scores of learning data, and Figure 3 shows an example of the data quintuple ratio according to different quantiles.

Referring to FIGS. 2 and 3 , a processor (eg, processor 200 of FIG. 1 ) may calculate an anomaly score based on an anomaly rate of data.

The processor 200 may configure a loss function using aggressive rejection and/or soft rejection in the process of training the neural network. At this time, ambiguous data that crosses a normal or abnormal distribution may be rejected.

The examples in FIGS. 2 and 3 may represent the anomaly score and anomaly ratio of an autoencoder learned using the MNIST dataset.

The example of FIG. 2 may represent the distribution of abnormal scores of training data in which 10% of the data is contaminated. Although the 0.9 quantile samples contain more abnormal data than the normal data, it can be seen that 6.7% of the data excluding the 0.9 quantile samples still contain abnormal data. In the case of the 0.5 quantile, there is more normal data than the 0.9 quantile, but the abnormality rate of the data excluding this may be about 0.034 (3.4%).

The processor 200 may perform aggressive rejection on data. Since the previous approach set the contamination ratio to be approximately 10%, 10% of the data can be treated as an abnormality. However, because normal and abnormal distributions overlap, it may be difficult to cover all anomalies. Additionally, assumptions about contamination rates can limit performance when contamination rates exceed 10%.

Based on this concept, aggressive rejection can remove anomalies while sacrificing a significant amount of normal samples. Figures 2 and 3 may show the abnormality score distribution of 10% contaminated data and the abnormality ratio for different quantiles. Figure 2 can show that normal and abnormal distributions overlap. Therefore, for more data, the contamination rate can be reduced as shown in Figure 3.

An outlier detection device according to an embodiment can improve the accuracy of outlier detection by utilizing a neural network with improved learning accuracy through a loss function to which a rejection method is applied, which will be described below. Equation 1 represents the loss function to which aggressive rejection is applied.

here,

represents the training data,

represents the weight for aggressive rejection,

represents the ith ideal score,

represents the q-displacement difference,

Can represent a model (e.g. autoencoder). The processor 200 determines the abnormal score through aggressive rejection.

The impact of larger data can be reduced. At this time, the ideal score can be defined for each model. For example, the anomaly score can be calculated based on the reconstruction error (e.g.

may be calculated based on . As shown in Figure 3, q may increase monotonically from 0.5.

The processor 200 may perform soft rejection on training data. Although aggressive rejection can reduce the contamination rate, it can reduce performance because it removes a large amount of normal data. Aggressive rejection can cause degradation of clean datasets. To solve this, the rejection weight

A second hyperparameter

It can be partially adjusted. In equation 2

at

Is

can depend on The weight can be expressed as Equation 2.

Processor 200 is

perform hard rejection to completely exclude the rejection target,

By performing soft rejection, learning can be partially performed using samples of the rejection target.

low

may exhibit low performance for clean datasets, and may exhibit low performance for clean datasets.

may reduce robustness, so the processor 200

You can set the value. For example, the model is

When is 0.1, minimal loss can be achieved for clean datasets and robustness can be achieved for polluted datasets.

FIG. 4 shows an example of a learning algorithm for a neural network used in the anomaly detection device of FIG. 1.

Referring to FIG. 4, a processor (eg, processor 200 of FIG. 1) may configure a heterogeneous loss function using rejection. The samples have an ideal score s of the neural network model.

If larger, they can be distinguished as potential abnormal samples.

About normal samples

can be set to 1, and for the remaining samples

It can be set to .

The ideal score of a mini batch is the gradient parameter

can be decided. The anomaly score is the modified z-score (or z-score) and input

based on

can be converted to

Neural network model parameters are soft rejection weights

and loss function

It can be updated using the loss generated by the product of .

The heterogeneous loss function used by the processor 200 can satisfy two conditions for using rejection.

As a first condition, the heterogeneous loss function may not minimize the abnormal loss function. This is because when a neural network model is trained to minimize loss, normal samples of the rejection target may interfere with learning.

As a second condition, the neural network model can be trained to converge quickly for normal samples, and the neural network model can be trained to converge more slowly than normal samples.

Processor 200 may determine a heterogeneous loss function to maximize the difference between the distribution of normal data and the distribution of abnormal data by adjusting the gradient.

As can be seen in Figure 2, the number of normal samples may increase as the abnormality score decreases. In order to perform more learning on normal samples and less learning on abnormal samples, the processor 200 may adjust the gradient based on the anomaly score.

The processor 200 may treat abnormal scores greater than the q quantile as potential abnormal samples, and treat the remaining samples as normal samples to apply a Mean Squared Error (MSE) loss. If the anomaly score is close to 0.5 quantile, the heterogeneous loss function can be changed to MSE. In the opposite case, the processor 200 may lower the gradient value.

The processor 200 may be configured using LGA, a dynamic gradient loss function of Equation 3, as a heterogeneous loss function. LGA is a parameter

It can be a loss function that handles from MSE loss to Welsch loss.

For example, the first loss function is in Equation 3:

can represent the case where is 2, and the second loss function is in Equation 3

It can indicate a case where is 1 or a value other than that.

Anomaly Score Gradient Parameter

To project to , processor 200 may use the z-score. The z score can be calculated as Equation 4,

represents the training data,

may mean the average of the training data, N may mean the number of training data, and c may mean the first hyperparameter.

The processor 200 may calculate the z-score based on the median of the data. The processor 200 may calculate the Median Absolute Deviation (MAD) of the data based on the median value of the data. Processor 200 may calculate the z-score based on the MAD.

Because the z-score uses the median, it can be robust to outliers. The z scores produce a normal distribution and can have relative distances. The z score can be normalized by the maximum of 3.5 and max(|z|), where 3.5 can be the threshold for outliers.

The processor 200 may define the value m using max(|z|). Normalized scores range from 0 to 1, and variance can take on smaller values. 3.5 can be used instead of a lower max(|z|) value to make the normalized z value closer to 0 when the variance is low.

3.5 used as a boundary value can accelerate convergence. Normalized z is the gradient parameter according to Equation 5

can be converted to

here,

represents the ideal score,

may represent the z score.

It may refer to a parameter for the lowest gradient loss.

The minimum value of can be matched to a normal loss function (MSE).

The maximum value of is a robust loss function, e.g.

) may be close to .

of

It can range from to 2.

Since the abnormality ratio increases in the form of a quadratic function as shown in the example of FIG. 3, z in Equation 5 can be applied in the form of a quadratic function.

Figure 5 shows an example of a graph for explaining a heterogeneous loss function, and Figure 6 shows another example of a graph for explaining a heterogeneous loss function.

Referring to Figures 5 and 6, the graphs of Figures 5 and 6 are

ego,

A heterogeneous loss function can be expressed using soft rejection. In Figure 5

It can be seen that as becomes smaller, the gradient decreases.

Figure 6

ego,

An example of a heterogeneous loss function can be shown in the case of . For potential outlier samples, a loss function between MSE and psuedo-Huber loss can be used, as shown in the shaded area of Figure 6.

For the lowest outliers, such as those with Z values close to 0 (e.g. 0.5 quantile), MSE with soft rejection can be used. For the largest outliers, which are the tails of the Z distribution, the psuedo-Huber loss using soft rejection can be used.

Figure 7 is a table showing AUROC (Area Under the Receiver Operating Characteristic) according to various loss functions, Figures 8 to 10 show an example of AUROC for the contamination rate of a plurality of models, and Figures 11 to 13 show a plurality of models. Shows another example of AUROC for the pollution rate of models, and Figures 14 to 16 show another example of AUROC for the pollution rate of a plurality of models.

Referring to FIGS. 7 to 16, the performance can be compared between a neural network model using an existing loss function and a neural network model using a heterogeneous loss function using rejection.

MNIST, F-MNIST (Fashion MNIST), and CIFAR-10 can be used as datasets. MNIST and F-MNIST can be composed of 10 classes and 28*28 scale images. CIFAR-10 can consist of 32*32 color images of 10 classes.

In an experiment, one class may be set as normal and the remaining classes may be set as abnormal. For normal data, the training data can be twice the test data, and 10% of the original training data can be used for validation.

Abnormal data may be added. here,

means the contamination rate, and N may mean the number of normal data. 30% of test data may consist of abnormal data.

Among neural network models, the model with the lowest verification loss can be used as a test model. Verification loss can be measured by aggressive rejection and hard rejection. The unit of performance may be AUROC (Area Under Receiver Operating Characteristic). In the experiment, each class is set to normal, and the average AUROC using three different seeds can be measured.

ITSR can be adopted as a neural network model to which RVAE-ABFA and NCAE are robust. ITSR can use OC-SVM and AAE for refinement. RVAE-ABFA can refer to a network that develops DAGMM by adopting attention based on VAE and feature adaptation. NCAE can use normal samples generated from a generative adversarial model to refine the dataset.

Loss functions can be evaluated based on three existing neural network models (eg AE, MemAE and DSVDD). DSVDD uses MSE, but the reconstruction error may be different. MemAE can be compared because of an additional loss function for memory augmented loss. MSE can be replaced by pseudo-Huber loss, GA loss and LOE loss. The GA loss uses the z distribution, whereas the heterogeneous loss function uses the z distribution.

Since negative log-likelihood is used to determine , it can be used to explain how GA loss affects robustness. In the table of FIG. 7, the heterogeneous loss function can be expressed as Hetero.

For MNIST and F-MNIST, the encoder may contain three convolutions. The decoder may include three deconvolutions. Except the last deconvolution, other deconvolutions or convolutions may include batch normalization, leaky Rectified Linear Unit (ReLU).

For CIFAR-10 data, the encoder and decoder may include four layers similar to the configuration of MNIST with different kernel and stride sizes. For the remaining models, a LeNET-based convolutional neural network with 32 or 128 representation dimensions can be used.

RVAE-ABFA can only use 32 dimensions due to computational errors. Batch size and epoch can be set to 100 and 300 respectively except for ITSR, NCAE and DSVDD. ITSR can refine data every 10 epochs after the first 100 epochs, and can be trained on the refined data using 100 epochs.

Parameters can be updated by the Adam optimizer using a learning rate of 0.0001 and a weight decay of 10^-6. Parameters in heterogeneous loss

,

and c are 0.5, 1.5 and

It can be.

To evaluate heterogeneous loss functions, robust loss functions can be used. Figure 7 shows the clean dataset (

=0) and contaminated dataset (

=0.2) can represent the AUROC. The highest AUROC is bolded.

As a result of comparison on clean datasets, the compared methods can show minimal AUROC loss compared to MSE, except for AE using MNIST. Heterogeneous loss functions on clean datasets can exhibit similar performance within 0.01 compared to MSE. The performance of AE using pseudo-Huber, GA, and LOE may not significantly improve robustness. However, it can be confirmed that AE using a heterogeneous loss function achieves a robustness of 0.045. When DSVDD for F-MNIST is learned using heterogeneous loss, it can produce the most effective results. LOE can exhibit performance similar to that of a heterogeneous loss function by maximizing an abnormal loss function. Because CIFAR-10 is a hard dataset compared to MNIST or F-MNIST, the AUROC for clean datasets may be lower. Therefore, robustness may not be significantly improved. The heterogeneous loss function can be more robust to 20% contaminated data at the level of 0.002 and 0.084 compared to the overall loss functions.

Figures 8 to 16 may show the results of visualizing the AUROC of robust loss functions depending on the contamination rate. Heterogeneous loss functions can achieve the most robust results in the cases of AE and MemAE. In the case of DSVDD for F-MNIST and LOE, it can outperform the heterogeneous loss function when the contamination rate is below 10%;

For =0.2, the heterogeneous loss function can outperform LOE. LOE is

value is actual

The highest performance can be achieved when is equal to . LOE can achieve robustness by maximizing the anomaly loss function;

=0.1 assumption may limit robustness. Because DSVDD is not a model based on reconstruction error,

Except for the case of =0.2, MSE and GA losses can be shown to be robust to CIFA-10. Experiments can demonstrate that the heterogeneous loss function can be applied to different MSE-based AD models and exhibits high robustness and minimal AUROC loss for clean datasets.

Figure 17 is a table showing the AUROC according to various neural network models, and Figures 18 to 20 show examples of comparing the AUROC for the contamination rate of the neural network model of the plurality of models in Figure 1.

17 to 20, standard models using heterogeneous loss can be compared with a robust neural network model. The table in Figure 17 can show the AUROC comparison results for different contamination rates. Although NCAE may show lower reduction in AUROC for contaminated datasets compared to clean datasets, its performance may be lacking. ITSR and RVAE-ABFA can exhibit more robust performance compared to AE and MemAE. However, the heterogeneous loss function can improve robustness to about 0.05 for AE and MemAE compared to ITSR. DSVDD using a heterogeneous loss function can show similar results to RVAE-ABFA and can show the most robust results for MNIST. RVAE-ABFA can show the highest AUROC compared to other methods for F-MNIST, but DSVDD using a heterogeneous loss function can show 0.012 higher robustness. Heterogeneous loss functions can achieve the highest robustness for CIFAR-10. Experimental results can show that the heterogeneous loss function exhibits performance comparable to other robust neural network models. 18 to 20 can show the robustness of neural network models according to various contamination rates. For the MNIST and F-MNIST datasets, RVAE-ABFA shows the most robust performance, but it can be seen that the heterogeneous loss function shows the most similar performance. When the contamination rate increases, the difference in DSVDD can be reduced by using RVAE-ABFA and the heterogeneous loss function. For CIFAR-10, DSVDD outperforms RVAE-ABFA, and the heterogeneous loss function can make MemAE have similar performance to RVAE-ABFA. Compared to other robust neural network models, it can be seen that the heterogeneous loss function achieves high robustness without increasing inference time and modifying the structure.

Figure 21 shows the results of a removal study for the heterogeneous loss function, and Figure 22 is a graph to explain the sensitivity of the heterogeneous loss function and soft rejection.

Referring to Figures 21 and 22, the impact after each component is removed in the experiment can be measured. The performance of clean and contaminated data sets in MNIST can be expressed as the table in FIG. 21.

and

can be a key component in determining the impact of aggressive rejection using soft or hard rejection. The loss function or the value of the loss function is

It can be expressed as,

represents the loss function and

Is

Indicates the displacement position,

may represent the degree of rejection.

Hard rejection - MSE(0.5,0) can produce more robust results for contaminated data than standard loss - MSE(1,1). However, hard rejection on a clean dataset can result in significant performance degradation for AE (0.115), MemAE (0.134), and DSVDD (0.045). This may be because many normal samples are excluded and has a negative impact on clean datasets. Soft Rejection - MSE (0.5, 0.1) can compensate for the decline. This can result in AUROC comparable to MSE on clean datasets. By using some excluded normal samples for training, soft rejection can improve the robustness of AE, MemAE, and DSVDD by 0.045, 0.07, and 0.044, respectively.

A heterogeneous loss function that does not perform aggressive rejection may be referred to as Hetero(1,1) in FIG. 21. This shows better performance than the baseline MSE and can show more robust performance than MemAE using MSE(0.5, 0.1). In addition, Hetero(0.5, 0.1) using aggressive rejection can show performance improvements of 0.008 and 0.15 compared to MSE (0.5, 0.1) for AE and MemAE, respectively. Experiments can indicate that differences between normal and abnormal data may be caused by gradient adaptation based on mini-batch distributions.

Heterogeneous loss functions can make soft rejection less sensitive to ts. While the heterogeneous loss function is robust due to gradient adaptation, soft rejection

If =1, it can be replaced with a non-robust loss function MSE. The sensitivity of the heterogeneous loss function and MSE using soft rejection can be expressed as shown in Figure 22. Soft rejection is

It can be seen that the effect decreases as . However, in the case of the heterogeneous loss function, which is less sensitive than MSE, on 20% contaminated data

= 0.1 and

You can make the difference between =0.5 smaller.

Referring to FIG. 23, the receiver 100 may receive data to detect whether it is normal or abnormal (2310).

The processor 200 can detect whether the data is normal or abnormal by inputting the data received from the receiver 100 into a pre-trained neural network (2330). For example, the processor 200 can detect abnormalities in the data by inputting data into a neural network learned to minimize the heterogeneous loss function described above, and the heterogeneous loss function is Equation 3 described above. It may be a function calculated based on Equation 5.

Depending on the embodiment, the processor 200 may apply a loss function to which the rejection described above is applied in addition to a heterogeneous loss function to learn a neural network used for outlier detection, and the process of applying rejection is: It may be the same as described through Equation 1 to Equation 3 described above.

Referring to FIG. 24, the receiver 100 may receive training data (2410).

The processor 200 may determine a gradient parameter based on the anomaly score of the training data received from the receiver 100 (2430). The processor 200 may calculate an anomaly score based on the restoration error of the learning data.

The processor 200 may determine a heterogeneous loss function based on the gradient parameter (2450).

The processor 200 may train an artificial neural network that detects outliers based on a heterogeneous loss function (2470). For example, the processor 200 may train the artificial neural network in a way that minimizes the calculated heterogeneous loss function. Additionally, the heterogeneous loss function may be a function calculated based on Equations 3 to 5 described above.

In addition, the processor 200 can train an artificial neural network to detect outliers based on a loss function to which the rejection described in FIG. 4 is applied, and the rejection is one of the aggressive rejection or soft rejection described above. It will be understood by those skilled in the art that this can be implemented.

The device described above may be implemented with hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), etc. , may be implemented using one or more general-purpose or special-purpose computers, such as a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. A processing device may execute an operating system (OS) and one or more software applications that run on the operating system. Additionally, a processing device may access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, a single processing device may be described as being used; however, those skilled in the art will understand that a processing device includes multiple processing elements and/or multiple types of processing elements. It can be seen that it may include. For example, a processing device may include a plurality of processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are possible.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. Software and/or data may be used on any type of machine, component, physical device, virtual equipment, computer storage medium or device to be interpreted by or to provide instructions or data to a processing device. , or may be permanently or temporarily embodied in a transmitted signal wave. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the medium may be specially designed and configured for the embodiment or may be known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -Includes optical media (magneto-optical media) and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

As described above, although the embodiments have been described with limited examples and drawings, various modifications and variations can be made by those skilled in the art from the above description. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

Therefore, other implementations, other embodiments, and equivalents of the claims also fall within the scope of the claims described below.

Claims

In an anomaly detection device using a heterogeneous loss function,

A receiver that receives data to detect whether it is normal or abnormal; and

A processor that detects abnormalities in the data by inputting the data into a pre-trained neural network.

Including,

The pre-trained neural network is,

An anomaly detection device that is learned through a heterogeneous loss function calculated using a gradient parameter determined based on the anomaly score of the learning data.
According to paragraph 1,

The above abnormal score is,

Calculated based on the restoration error,

Anomaly detection device.
According to paragraph 1,

The gradient parameter is,

Calculated based on the z score calculated based on the median of the learning data and the maximum absolute value of the z score,

Anomaly detection device.
According to paragraph 3,

The z score is,

Calculated based on the MAD (Median Absolute Deviation) of the learning data, which is calculated based on the median of the learning data,

Anomaly detection device.
According to paragraph 3,

The gradient parameter is,

Determined based on the larger value of a predetermined constant and the maximum value,

Anomaly detection device.
According to paragraph 3,

The heterogeneous loss function is determined based on Equation 1 to Equation 3,

Equation 1 above is:

ego,

Equation 2 above is:

ego,

Equation 3 above is:

represents the training data,
is the average of the training data, N is the number of training data, c is the first hyperparameter,
represents the ideal score,
is the z score,
Anomaly detection device, which refers to the parameters for lowest gradient loss.
According to paragraph 1,

The above abnormal score is,

An abnormality detection device calculated based on the restoration error.
According to paragraph 1,

If the anomaly score is greater than a predetermined value, the rejection weight is determined as a first value,

If the abnormality score is less than or equal to a predetermined value, the rejection weight is determined as a second value,

Anomaly detection device.
According to clause 8,

The first number is,

It is determined as a value less than 0 or 1,

The second figure is,

determined by 1,

Anomaly detection device.
According to clause 8,

The loss function is determined based on Equation 4 to Equation 5,

Equation 4 above is:

ego,

Equation 2 above is:

ego,

represents the training data,
represents the rejection weight,
represents the q-displacement difference, L represents the loss function,
Indicates the model (e.g. autoencoder),
represents a hyperparameter, an anomaly detection device.
In a method of anomaly detection,

Receiving data to detect whether it is normal or abnormal; and

Detecting whether there is an abnormality in the data by inputting the data into a pre-trained neural network

Including,

The pre-trained neural network is,

Learned based on a heterogeneous loss function calculated through a gradient parameter determined based on the abnormality score of the learning data.

An anomaly detection method learned through.
In a learning method using a heterogeneous loss function,

Receiving training data; and

determining a gradient parameter based on the abnormality score of the learning data;

determining a heterogeneous loss function based on the gradient parameters; and

Learning an artificial neural network based on the heterogeneous loss function

Learning methods including.