A Survey on Applications of One Class Classiﬁcation

Navya V K and Shyam P Joy

Department of Artiﬁcial Intelligence and Machine Learning, CMR Institute of Technology,

Bengaluru, Karnataka 560037, India

Keywords:

OCC Survey, Class Imbalance, Anomaly Detection, Outlier Detection, Product Defect Detection, Fraudulent

Transaction.

Abstract:

One Class Classiﬁcation (OCC) is used to address the issues related to class imbalance datasets where there

are less samples of negative class, and training is done using a single positive class.Many algorithms have been

developed to generate OCC models. This paper presents a survey on applications using OCC.The surveys on

similar topics were covered from 2010 to 2021. This paper presents the research developments after 2021. The

various methods used with each application along with the different datasets and its accuracies are discussed

1 INTRODUCTION

The main aim of One Class Classiﬁcation is to

build the classiﬁcation models when the negative

class is poorly sampled, not well deﬁned or be ab-

sent. So, deﬁning a class boundary using the positive

class alone lead towards OCC. The term One Class

Classiﬁcation was coined by Moya et al. (Moya et al.,

1993) in their research work during 1993. The ability

to detect the abnormal objects, outliers or suspicious

patterns from the normal objects makes OCC more

popular.

Imbalanced classes pose more challenges com-

pared to balanced classes. (Hasanin et al., 2019).For

example the feature selection and achieving high ac-

curacy are more difﬁcult while using OCC.

The survey on applications of OCC is available

from 2010 till 2021. This survey presents the applica-

tions of OCC published after 2021. The applications

considered are classiﬁed into the following categories

like Anomaly Detection, Novelty Detection, Intrusion

Detection, Fraudulent Transaction and Product Defect

identiﬁcation. The methods were based on OCC and

its accuracies are given in the below sections.

The following section presents the surveys on ap-

plications of OCC published prior to 2021.

2 RELATED WORK

In 2010, Shehroz S Khan et al. (”Khan and Madden,

2010) had published a survey on trends in OCC.This

paper discussed the taxonomy, algorithms,the domain

of application used, and the overall review of OCC

work was discussed from 2006 to 2011.

In 2021, Pramuditha Perera et al.(Perera et al.,

2021) published a survey of their research during

2007-2020 which was on deep learning based OCC

for visual recognition.

In 2021, Naeem Seliya et al.(Seliya et al., 2021)

provided the survey of algorithms, approaches and

methodologies for OCC from 2010 -2021. The dif-

ferent works under OCC were surveyed in three cat-

egories like Outlier detection, Novelty detection and

Deep learning.

A number of techniques has been applied to

solve many problems.In the following section,we

present the applied research using OCC,classiﬁed into

various categories of applications.The categories of

applications are Anomaly Detection,Novelty Detec-

tion,Intrusion Detection,Fraudulent Transactions and

Product Defect Identiﬁcation.

3 APPLICATIONS

3.1 Anomaly Detection

Anomaly detection is also known as Outlier de-

tection and for a given dataset the anomaly detection

identiﬁes the uncharacteristic data sample (known as

outliers or anomalies). One class classiﬁcation (OCC)

concentrates more on the data from the positive class

and there is a learned classiﬁer which deﬁnes the

748

V K, N. and P Joy, S.

A Survey on Applications of One Class Classiﬁcation.

DOI: 10.5220/0013585100004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 1, pages 748-755

ISBN: 978-989-758-763-4

boundary between the positive and negative class.

OCC is used in the areas of Machine Learning, Com-

puter vision. So given below is the survey which

shows the developments of anomaly detection using

One class classiﬁcation during the years 2022 and

2023.

In 2022, Pauline Arregoces et al. (Arregoces et al.,

2022) concentrated on the Anomaly based Intrusion

Detection System where they try to identify the mali-

cious behaviour by modelling a system with normal

behaviour. They used UNSW-NB15 attack dataset

which consist of normal and abnormal trafﬁc. In ab-

normal trafﬁc they included Denial of Service attacks,

worms, shellcodes etc. The dataset was highly imbal-

anced as the normal trafﬁc was 87.35% and imbal-

anced trafﬁc was 12.65%. The One class SVM, Iso-

lation Forest were used with two encoding schemes

such as One hot encoding and Label encoding. These

encoding schemes were applied to full and reduced

dataset which was balanced and unbalanced and re-

ceived an AUC (Area under the receiver operating

characteristic Curve) of 98%.

J Pang et.al (Pang et al., 2022) used a hybrid

algorithm combining vector quantization (VQ) and

OCSVM (One class Support Vector Machine). Prior

to this OCSVM was a popular method for anomaly

detection but it relies on the Kernel parameters so for

uneven and complex distribution of data it may be dif-

ﬁcult to get the god boundaries. So, Vector quantiza-

tion obtains the distribution of normal data and map

data to a feature space that is high dimensional.The

classiﬁer carries the data, and an integration of gener-

ative and discriminant learning is done.

Renuka Sharma et al. (Sharma et al., 2022) pro-

posed a Deep Neural Network (DNN) framework

which is semi supervised variational learning which

leverages the generalized Gaussian model on the la-

tent space and reconstructed image which is used in

anomaly detection. So the ssgVAE(semi supervised

gaussian Variational Encoder) can leverage the outlier

data during training to improve the performance. It

can classify images into normal and abnormal classes.

It is modelled with an encoder and decoder. They

applied to real world datasets and gave a better per-

formance of AUC (Area under the receiver operating

characteristic Curve) in the range 0.80-0.89%.

Hongzuo Xu et al. (Xu et al., 2022a) proposed

a Calibrated One class classiﬁcation for Unsuper-

vised time series anomaly detection (COUTA) where

the classiﬁer is calibrated by discriminating the nor-

mal samples with the abnormal behaviour and iden-

tify the class of abnormal behaviour. They tried

to address the presence of anomaly contamination

and absence of knowledge about anomalies. They

used two methods Uncertainty Modelling based Cal-

ibration (UMC) which helps to exclude the contam-

inated data from the training based on the uncer-

tainty scores while Native Anomaly based Calibra-

tion (NAC) which helps in identifying the anomaly

behaviour by creating a proper normality bound-

ary.There were twelve anomaly method with which

COUTA was been compared. COUTA gave better

performance compared to all the methods as shown

in the below table 1.

Table 1: Performance of COUTA with six different datasets

Dataset AUC

ASD 0.955

SMD 0.984

SWaT 0.900

WaQ 0.714

DSADS 0.942

Epilepsy 0.823

In 2023, Mohanad Sarhan et al. (Sarhan et al.,

2023) proposed a DOC, deep one class classiﬁcation

for network anomaly identiﬁcation. They used a deep

Support Vector Data Description that can map net-

work data features to a low dimension embedding.

Later it is applied to HBOS, Histogram based Out-

lier Score which can identify the benign or an attack.

The two NIDS (Network Intrusion Detection System)

datasets and DOC classiﬁer gave an F1 score of 98.16

and AUC of 98.89 compared to the other classiﬁers

(PCA, Deep SVDD etc).

Rui wang et. al. (Wang et al., 2023b) sug-

gested a deep Contrastive One Class Anomaly detec-

tion (COCA) following the contrastive learning and

the OCC. A Sliding window is used to divide the time

series into T length sequences, where T is the length

of the sliding window. If d is the dimension, when

d = 1 it is univariate and when d > 1 it is multivari-

ate. If set of time series is given by D = X

, X

..X

the anomaly score S

is calculated for every X

. If

the S

is higher then it is a anomalous time series.

The AIops and UCR were the two datasets used. The

COCA model gave an F1 score of 66.78 for AIops

and for UCR an F1 score of 79.16. The accuracy of

this model was 66.12.

In 2023, Marcella Astrid et al. (Astrid et al., 2023)

proposed an anomaly reconstruction capability using

One class classiﬁer to detect the video anomalies. In

OCC, an Autoencoder is usually trained to recon-

struct the training data with normal samples, and it

was found to be poor in reconstructing the anoma-

lous data. To reconstruct normal as well as anoma-

lous data the Autoencoder is trained with both normal

and anomalous data. The pseudo anomalies are also

incorporated during the training of an Autoencoder.

A Survey on Applications of One Class Classiﬁcation

749

They have used ﬁve different methods like Pseudo

bound Skip Frames, Pseudo bound Repeat Frames,

Pseudo bound Patch, Pseudo bound Fusion, Pseudo

bound Noise(Astrid et al., 2023), to synthesize pseudo

anomalies for the anomaly detection. This was ap-

plied to Ped2, Avenue, ShanghaiTech datasets and

table 2 results was obtained where PSNR is Peak

Signal-to-Noise Ratio and MSE is Mean Squared Er-

ror.

3.2 Detection of Fraudulent

Transactions

In 2022, K.S.N.V. K Gangadhar et.al (Gangadhar

et al., 2022) proposed a One class classiﬁcation that

was based on Chaotic Variational Autoencoder(C-

VAE). It is used in Insurance fraud detection. There

will be an encoder and a decoder in both Variational

Autoencoder (VAE) and C-VAE. The difference be-

tween VAE and C-VAE is in the latent distribution

component which is used to generate the chaotic map

(Gangadhar et al., 2022). The logistic chaotic map is

incorporated with C-VAE which are helpful in feature

subset selection (Vivek et al., ) and in handling imbal-

ance dataset (Vivek et al., ). This C-VAE is applied to

health care dataset and in non-health care dataset like

automobile insurance dataset. The classiﬁcation rate

of VAE based One class Classiﬁcation is compared

with C-VAE based One class classiﬁcation and as we

could see C-VAE gives better results compared to the

VAE as shown in the below table 3.

Table 3: Mean classiﬁcation rate of VAE and C-VAE ap-

plied to Medicare and Automobile Insurance datasets

Dataset Model MCR

Medicare VAE 73.13

Medicare C-VAE 77.9

Automobile insurance VAE 86.9

Automobile insurance C-VAE 87.25

*MCR Mean classiﬁcation Rate

In 2022, Yellati Vivek et al. (Vivek et al., 2022)

proposed a fraud detection which based on binary

classiﬁcation or one class classiﬁcations. In the fraud

detection framework they incorporated an Explain-

able Artiﬁcial Intelligence(XAI) and a Casual Infer-

ence (CI). In binary classiﬁcation they used over-

sampling techniques like Generative Adversarial Net-

works (GAN) and Synthetic Minority Oversampling

Technique (SMOTE).

In OCC, the model is built on negative class, and it

was tested on the positive class. They used a train and

split mechanism which helps to normalize the nega-

tive class data which builds the OCC model. The dif-

ferent OCC algorithms (Vivek et al., 2022) were ap-

plied to ATM transaction dataset and Isolation Forest

(IForest) gave the better classiﬁcation rate. The dif-

ferent OCC algorithms with their classiﬁcation rate

results are given in the below table 4.

Table 4: Classiﬁcation rate of different OCC models ap-

plied to ATM Transactions dataset

OCC Model CR

OCSVM (One-class SVM) 0.947

IForest (Isolation Forest) 0.959

COPOD (Copula-Based Outlier Detection) 0.760

Angle-based Outlier Detection (ABOD) 0.941

MCD (Minimum Covariance Determinant) 0.943

VAE (Variational Autoencoder) 0.547

*CR Classiﬁcation Rate

In 2023, Zaffer et al. (Zaffar et al., 2023) pro-

posed a sub space learning approach which was based

on One class classiﬁcation. It could handle the im-

balance data and automatically detect the fraudu-

lent transactions based on credit card. They used

a Subspace Support Vector Data Description model

(SSVDD) where the data is projected using a pro-

jection matrix after reducing the dimensional sub-

space and it incorporated geometric information to

ﬁnd the optimized set of features in Graph Embedded

SSVDD(GESSVDD). They used four datasets (Zaf-

far et al., 2023) which were imbalanced, and these

datasets were based on credit card transactions, digital

payment transactions, mobile transactions, and bank

transactions. They applied various models like One

class SVM, OCSVM, SVDD, SSVDD, GESSVDD-

kNN etc. The G-means was calculated which is used

to show the model accuracy. The GESVDD with

kNN (k-Nearest Neighbours) gave the better results

as shown in the below table5.

Table 5: G-means of four datasets which was applied on

different OCC models

Dataset G-means

D1 0.906

D2 0.692

D3 0.728

D4 0.691

3.3 Intrusion Detection

In 2022, Wen Xu et al. (Xu et al., 2022b) proposed a

Bidirectional GAN(Generative Adversarial Network)

which can be used for intrusion detection and to re-

duce the training overhead. In GAN there will be a

network generator and a discriminator. The gener-

ator can produce the output same as the input data

while a discriminator takes fake data and real data as

inputs and distinguish them. There were two issues

INCOFT 2025 - International Conference on Futuristic Technology

750

Table 2: Performance of ﬁve different methods of pseudo anomalies with Ped2, Avenue, ShanghaiTech datasets

Ped2 Avenue ShanghaiTech

PSNR MSE PSNR MSE PSNR MSE

PseudoBound-

Skip frames

98.44% 98.39% 87.10% 84.37% 73.66% 72.31%

PseudoBound-

Repeat frames

93.69% 94.20% 81.87% 80.71% 72.58% 71.64%

PseudoBound-

Patch

95.33% 95.15% 85.36% 78.26% 72.77% 72.03%

PseudoBound-

Fusion

94.16% 94.34% 82.79% 81.60% 71.52% 70.97%

PseudoBound-

Noise

97.78% 97.31% 82.11% 79.63% 72.02% 71.69%

while we use a GAN 1) There is a strong dependency

that could exist between the generator and discrimi-

nator and 2) The complexity involved in generating

anomaly scores to segregate the input samples to nor-

mal or not. BIGAN (Bidirectional GAN) is a variant

of GAN where an encoder is being added. The depen-

dence of the data can be reduced by training the gener-

ator and encoder until they produce new data samples

that could resemble the original data. The construc-

tion of one class classiﬁer for detecting normal trafﬁc

with abnormal trafﬁc rather than for calculating the

anomaly scores are normally complex and expensive.

The table6 given below shows the data sets used

and the F1-score after applying the BIGAN,

Table 6: F1-score of datasets when BIGAN was applied

S.no Data Set used F1-score

1 NSL-KDD 92%

2 CIC-DDoS2019 99%

In 2022, Esteba et al. (Jove et al., 2022) imple-

mented an Intrusion Detection System based on a dif-

ferent one class classiﬁers with the motive of pre-

venting attacks over the IoT using Message Teleme-

try Transport (MQTT) as case study. MQTT is a

messaging protocol, and it is designed for light ma-

chine to machine communication. It is normally used

to connect small devices to a network with limited

bandwidth. This paper mainly concentrates on im-

proving the security in IoT environment by present-

ing certain characteristics which makes a target of

attack. An intrusion detection classiﬁer is built by

training the dataset without any attacks. So later if

we apply any data set with or without attacks this

one class classiﬁer makes the prediction more accu-

rately. The anomaly methods like Convex hull, non-

convex boundary method, K-means, Principal Com-

ponent analysis(PCA) were used and in PCA got the

highest accuracy of 89%.

In 2023, Wenbin Yao et al. (Alazzam et al., 2022)

has worked on detecting unknown attacks. The un-

known attacks normally security expert needs to con-

ﬁrm, and retraining is needed for the model to make

the correct predictions. So, a Lightweight Intelligent

Network-based intrusion detection system (NIDS)

was proposed which used a Bidirectional Gated Re-

current Unit (BGRU) Autoencoder and an Ensemble

learning. BGRU network learns the data and it is ap-

plied to the model where the normal data returns a

loss with small range but abnormal data returns with

big loss. The model has high prediction accuracy

even if it was trained with normal data. The ensem-

ble learning uses random forest, Light GBM (Gradi-

ent Boosting Machine), XGBoost as the base classi-

ﬁers and a soft voting techniques is used to predict the

result of the three classiﬁers for the optimal classiﬁca-

tion. This BGRU with ensemble learning was applied

to the datasets and the following table 7accuracy was

obtained.

Table 7: F1-score of datasets when BGRU method was

applied

S.no Data Set used F1-score

1 WSN-DS 97.91%

2 UNSW-NB15 98.92%

3 KDD CUP99 98.23%

3.4 Novelty Detection

In Novelty detection, the test data can be classiﬁed

as normal or novel using the training data even if, no

novel instances exist in the training data. One class

classiﬁcation uses some novel detection approaches to

do this. So given below is the survey which shows the

developments of Novelty detection using One class

classiﬁcation during the years 2022 and 2023.

In 2022, Shao Yuan Lo et al. (Lo et al., 2022)

proposed an Adversarial robust, one class novelty

detection method. An adversarial attack are some

specialized inputs which can fool a deep neural net-

work leading to misclassiﬁcation and these types of

A Survey on Applications of One Class Classiﬁcation

751

attacks were not properly investigated in the context

of one class novelty detection. So, the basic idea

is to train a novelty detector where the latent space

can be manipulated called as Principal latent space

where the principal components learn incrementally

and thus improves the adversarial robustness. This

Principal latent space reconstructs both normal and

novel class into known classes and novel classes al-

ways has high reconstruction errors compared to the

normal data. The principal latent space was applied to

three datasets with seven novelty detection methods to

identify eight adversarial attacks and a consistent im-

provement against adversarial robustness were found

as shown in the below table 8.

Table 8: Accuracy of the datasets MNIST,F-

MNIST,CIFAR-10 when PrincipalLS was applied

Defense MNIST F-MNIST CIFAR-10

PrincipaLS 0.973 0.922 0.578

John Taylor et al. (Jewell et al., 2022) suggested

a Learned Encoder-Decoder network based on One

class where an adversarial context masking is done

for the novelty detection. Generally, Autoencoders

are used for the novelty detection and there will be an

encoder to map the image to a lower dimension latent

space and decoder to map from the latent space to the

original space. The context autoencoders are more

effective since they reconstruct the images from the

masked images, but it can obtain only a suboptimal

representation. Here, the authors have suggested two

modules i.e., Mask module which learns to generate

the masked images and a Reconstructor to reconstruct

the masked images to get the original image. It was

applied to datasets like MNIST, CIFAR10, UCSD and

99.02% AUC was obtained.

In 2022, Joshua L et al. (Pulsipher et al., 2022)

proposed a feature extraction based on sensor using

the One class classiﬁcation. They have made use

of computer vison sensors. The Convolutional Neu-

ral Network is a supervised learning models used in

Computer vison where the data can be extracted from

an image or visual data. So, there will be an extractor

where the input images are fed to specialized convolu-

tional layers to extract the visual patterns. The predic-

tor is fed with the extracted features, to get the desired

output states. The identiﬁcation of novelty in real time

is difﬁcult so a SAFE-OCC (Sensor Activated Feature

Extraction OCC) was proposed where a CNN model

maps the visual data to a state signal and this signal

will be interpreted with the help of a controller with

automated control loops to extract information. This

approach gave best results for the various forms of

data ranging from 96-100%.

In 2023, Biao Wang et al. (Wang et al., 2023a)

proposed an ensemble detector where they used Se-

lective Feature Bagging which is an improved version

of Feature Bagging. It is a dynamic classiﬁer selec-

tion and in ensemble generation phase the base detec-

tors are divided into various groups with space dimen-

sion. It enhances the accuracy through considering

the data variance and bias that occurs within the data.

It can be used with high dimension data and dynamic

selection helps in the bias reduction.

3.5 Product Defect Detection

In 2023, Seunghun Lee et al. had proposed a defect

inspection method based on One class Classiﬁcation

to deal with the imbalanced datasets. This overcomes

the Representation collapse problem.Here, the train-

ing data follows a repetitive pattern, or the training

data diversity is insufﬁcient which could result in per-

formance degradation. So, a two-stream network One

class Classiﬁcation was implemented.

The two-stream network has global and local fea-

ture extraction network. In global feature extraction

network, they learn the general features of the target

class and local feature extractor is designed to capture

the speciﬁc feature from the training dataset. The fea-

ture vector output from each network will be merged

and it will send through the classiﬁcation layer to

make the ﬁnal prediction. This model was applied

to an image dataset consisting of automotive airbag

bracket inspection and obtained an F1 score of 93%.

In 2023, Chaabi et al. (Chaabi et al., 2023) sug-

gested a solution towards automatic defect detection

using One Class Classiﬁcation which is to improve

the quality of a control system. Normally a visual in-

spection has been done to identify the defects, but it

can be error prone, and another issue is of class imbal-

ance where the normal samples are readily available

but a very few or none of the defected samples are

available.

The product surface images are used as input and

only images of the normal class are present during

training. They have used three sub models Convo-

lutional Autoencoder (CAE), Principal Component

Analysis (PCA), Support Vector Data Description

(SVDD). The Convolutional autoencoder which ex-

tracts abstract features from the images and the ex-

tracted feature vector is fed to Principal Component

Analysis (PCA) for dimensionality reduction. The

reduced dimension data is fed into one class classi-

ﬁer, SVDD. During the test phase both normal and

defected images are being used. It was applied to

MVTec-AD (Anomaly Detection) dataset were used

and using CAE-PCA-SVDD they obtained an F1

score of 97%.

INCOFT 2025 - International Conference on Futuristic Technology

752

In 2023, Sang-Min Kim et al. (Kim and Sohn,

2023) proposed a one class-based vibration anomaly

detection for diagnosing the defects. The character-

istics of vibration data is extracted using an Autoen-

coder through compression and restoration process.

The frequency characteristics of the vibration data is

not considered.A Convolutional Autoencoder is used

to extract features and the anomaly detection is per-

formed with the help of one class classiﬁer. So, they

proposed a dilated Convolution which adds dilation

to the general convolution by expanding the gap be-

tween kernels through which many input values are

being referred to obtain one sample value. There is a

Multicolumn autoencoder which focuses on extract-

ing the features of vibration data by reﬂecting the fre-

quency characteristics. The table 9 given below com-

pares various models with their accuracy and Multi-

column Autoencoder shows the better results.

Table 9: Accuracy of the models applied to datasets based

on one class based vibration anomaly detection

Model Accuracy

Convolutional Autoencoder 0.734

LSTM Autoencoder 0.885

Multicolumn Autoencoder 0.910

3.6 Medical Image Classiﬁcation

In 2018, Qi Wei et al. (Wei et al., 2018) proposed

an unsupervised model using which we can charac-

terize the negative class. To characterize the negative

patches, they used an autoencoder based on deep neu-

ral network. A testing image was decomposed into

patches and autoencoder reconstructs these patches

and classify it into positive or negative one which ﬁ-

nally leads to a one class classiﬁer. So, in the medi-

cal image the positive patches show the suspicious ar-

eas which contain the anomalies. They applied this

method on Breast dataset for the disease detection

and obtained an Area under the ROC Curve (AUC)

of 0.84.

In 2019, Yu-Xing Tang et al. (Tang et al., 2019)

suggested a one class learning which was based on

Deep Adversarial for classifying the normal and ab-

normal chest radiograph. They used end to end, semi

supervised DAOL (Deep Adversarial One class learn-

ing). They train the system by taking normal X-ray

images and it is used to classify normal and abnormal

chest radiograph X-ray classiﬁcation. They applied

this to NIH chest X-ray dataset and obtained an AUC

of 0.80.

In 2020, Long Gao et al. (Gao et al., 2020) worked

on how to handle unbalanced medical image data us-

ing one class classiﬁcation with deep learning. Prior

to this, the one class modelling was restricted to med-

ical images with complex features. So, they applied

Image Complexity based One Class Classiﬁcation

(ICOCC) and they implemented using the Convolu-

tional Neural Network. Here the basic idea was to dis-

tinguish between the perturbed and original images.

The classiﬁer learns the discriminative features during

the training, and it helps it to distinguish the samples

of other perturbed image class. They used the MRI,

FFDM, SOKL and Hep-2 data set (Gao et al., 2020)

and achieved the following F1-score which measures

the model accuracy as shown in the below table 10.

Table 10: F1-Score obtained by the datasets when applied

with ICOCC

Dataset F1score

MRI .969

FFDM .924

SOKL .703

Hep-2 .941

In 2022, Eduardo Perez et al (Perez-Careta et al.,

2022) used OCC to classify chest radiographs to iden-

tify the presence of COVID-19. OCC is helpful in

identifying the diseases from the images in the early

stage diagnosis.They used two datasets one with im-

ages from the patients with COVID-19 and X-ray

images of the patients with active respiratory condi-

tion. They used the X-ray images of 126 which are of

COVID -19 infected, and 165 normal X-ray images

and they were applied on RBF (Radial basis func-

tion), Linear and Isolated Forest. In this Isolated For-

est gave the better results with an F1-Measure of 0.61

without enhancement and 0.63 with enhancement.

4 CONCLUSION

One class classiﬁcation is a special case of multi

class classiﬁcation.In binary classiﬁcationthe samples

of positive and negative classes are required for the

classiﬁcation.But,OCC could be applied to the appli-

cations where the samples of one class is available

but, there is a difﬁculty in getting the samples of other

class.OCC has the capability to learn the character-

isitcs even if the data contains noise or errors. This

survey presents a categorization of various applied

research using OCC during the year 2022-23 and it

would help the researchers who work with problems

of Class imbalance, which can be solved using OCC.

REFERENCES

Alazzam, H., Sharieh, A., and Sabri, K. E. (2022). A

lightweight intelligent network intrusion detection

A Survey on Applications of One Class Classiﬁcation

753

system using ocsvm and pigeon inspired optimizer.

Applied Intelligence, 52(4):3527–3544.

Arregoces, P., Vergara, J., Guti

errez, S. A., and Botero,

J. F. (2022). Network-based intrusion detection: A

one-class classiﬁcation approach. In NOMS 2022-

2022 IEEE/IFIP Network Operations and Manage-

ment Symposium, pages 1–6. IEEE.

Astrid, M., Zaheer, M. Z., and Lee, S.-I. (2023). Pseu-

dobound: Limiting the anomaly reconstruction capa-

bility of one-class classiﬁers using pseudo anomalies.

Neurocomputing, 534:147–160.

Chaabi, M., Hamlich, M., and Garouani, M. (2023). Prod-

uct defect detection based on convolutional autoen-

coder and one-class classiﬁcation. IAES International

Journal of Artiﬁcial Intelligence, 12(2):912.

Gangadhar, K. S. N. V. K., Kumar, B. A., Vivek, Y., and

Ravi, V. (2022). Chaotic variational auto encoder

based one class classiﬁer for insurance fraud detec-

tion.

Gao, L., Zhang, L., Liu, C., and Wu, S. (2020). Handling

imbalanced medical image data: A deep-learning-

based one-class classiﬁcation approach. Artiﬁcial in-

telligence in medicine, 108:101935.

Hasanin, T., Khoshgoftaar, T. M., Leevy, J. L., and Bauder,

R. A. (2019). Severely imbalanced big data chal-

lenges: investigating data sampling approaches. Jour-

nal of Big Data, 6(1):1–25.

Jewell, J. T., Khazaie, V. R., and Mohsenzadeh, Y. (2022).

One-class learned encoder-decoder network with ad-

versarial context masking for novelty detection. In

Proceedings of the IEEE/CVF winter conference on

applications of computer vision, pages 3591–3601.

Jove, E., Aveleira-Mata, J., Alaiz-Moret

on, H., Casteleiro-

Roca, J.-L., Marcos del Blanco, D. Y., Zayas-Gato,

F., Quinti

an, H., and Calvo-Rolle, J. L. (2022). Intel-

ligent one-class classiﬁers for the development of an

intrusion detection system: the mqtt case study. Elec-

tronics, 11(3):422.

”Khan, S. S. and Madden, M. G. (”2010”). ”a survey of

recent trends in one class classiﬁcation”. In ”Coyle,

L. and Freyne, J., editors, ”Artiﬁcial Intelligence and

Cognitive Science”, pages ”188–197”, ”Berlin, Hei-

delberg”. ”Springer Berlin Heidelberg”.

Kim, S.-M. and Sohn, J.-M. (2023). Vibration anomaly de-

tection of one-class classiﬁcation using multi-column

autoencoder. Journal of The Korea Society of Com-

puter and Information, 28(2):9–17.

Lo, S.-Y., Oza, P., and Patel, V. M. (2022). Adversarially

robust one-class novelty detection. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

45(4):4167–4179.

Moya, M. M., Koch, M. W., and Hostetler, L. D. (1993).

One-class classiﬁer networks for target recognition

applications. NASA STI/Recon Technical Report N,

93:24043.

Pang, J., Pu, X., and Li, C. (2022). A hybrid algo-

rithm incorporating vector quantization and one-class

support vector machine for industrial anomaly detec-

tion. IEEE Transactions on Industrial Informatics,

18(12):8786–8796.

Perera, P., Oza, P., and Patel, V. M. (2021). One-class classi-

ﬁcation: A survey. arXiv preprint arXiv:2101.03064.

Perez-Careta, E., Hern

andez-Far

ıas, D. I., Guzman-

Sepulveda, J. R., Cisneros, M. T., Cordoba-Fraga, T.,

Martinez Espinoza, J. C., and Guzman-Cabrera, R.

(2022). One-class classiﬁcation for identifying covid-

19 in x-ray images. Programming and Computer Soft-

ware, 48(4):235–242.

Pulsipher, J. L., Coutinho, L. D., Soderstrom, T. A., and

Zavala, V. M. (2022). Safe-occ: A novelty detection

framework for convolutional neural network sensors

and its application in process control. Journal of Pro-

cess Control, 117:78–97.

Sarhan, M., Kulatilleke, G., Lo, W. W., Layeghy, S., and

Portmann, M. (2023). Doc-nad: A hybrid deep one-

class classiﬁer for network anomaly detection. In

2023 IEEE/ACM 23rd International Symposium on

Cluster, Cloud and Internet Computing Workshops

(CCGridW), pages 1–7. IEEE.

Seliya, N., Abdollah Zadeh, A., and Khoshgoftaar, T. M.

(2021). A literature review on one-class classiﬁcation

and its potential applications in big data. Journal of

Big Data, 8:1–31.

Sharma, R., Mashkaria, S., and Awate, S. P. (2022). A semi-

supervised generalized vae framework for abnormal-

ity detection using one-class classiﬁcation. In Pro-

ceedings of the IEEE/CVF Winter Conference on Ap-

plications of Computer Vision, pages 595–603.

Tang, Y.-X., Tang, Y.-B., Han, M., Xiao, J., and Summers,

R. M. (2019). Deep adversarial one-class learning for

normal and abnormal chest radiograph classiﬁcation.

In Medical Imaging 2019: Computer-Aided Diagno-

sis, volume 10950, pages 305–311. SPIE.

Vivek, Y., Ravi, V., Mane, A. A., and Naidu, L. R. (2022).

Explainable artiﬁcial intelligence and causal inference

based atm fraud detection.

Vivek, Y., Vadlamani, D. R., and Radha Krishna, P. Feature

subset selection for big data via parallel chaotic bi-

nary differential evolution island model under apache

spark. Available at SSRN 4133444.

Wang, B., Wang, W., Meng, G., Meng, T., Song, B., Wang,

Y., Guo, Y., Qiao, Z., and Mao, Z. (2023a). Selective

feature bagging of one-class classiﬁers for novelty de-

tection in high-dimensional data. Engineering Appli-

cations of Artiﬁcial Intelligence, 120:105825.

Wang, R., Liu, C., Mou, X., Gao, K., Guo, X., Liu, P., Wo,

T., and Liu, X. (2023b). Deep contrastive one-class

time series anomaly detection. In Proceedings of the

2023 SIAM International Conference on Data Mining

(SDM), pages 694–702. SIAM.

Wei, Q., Ren, Y., Hou, R., Shi, B., Lo, J. Y., and Carin,

L. (2018). Anomaly detection for medical images

based on a one-class classiﬁcation. In Medical Imag-

ing 2018: Computer-Aided Diagnosis, volume 10575,

pages 375–380. SPIE.

Xu, H., Wang, Y., Jian, S., Liao, Q., Wang, Y., and Pang, G.

(2022a). Calibrated one-class classiﬁcation for unsu-

pervised time series anomaly detection. arXiv preprint

arXiv:2207.12201.

INCOFT 2025 - International Conference on Futuristic Technology

754

Xu, W., Jang-Jaccard, J., Liu, T., Sabrina, F., and Kwak, J.

(2022b). Improved bidirectional gan-based approach

for network intrusion detection using one-class classi-

ﬁer. Computers, 11(6):85.

Zaffar, Z., Sohrab, F., Kanniainen, J., and Gabbouj, M.

(2023). Credit card fraud detection with subspace

learning-based one-class classiﬁcation.

A Survey on Applications of One Class Classiﬁcation

755