Detection of Malicious Applications using YOLO V3-Spatial Pyramid

Pooling over Optical Character Recognition for

Computing Access Time

Gowtham V. and Devi T.

Saveetha University, India

Keywords: Android Application, Deep Learning, Novel You Only Look Once V3-Spatial Pyramid Pooling, Internet

Access, National Security, Optical Character Recognition Algorithm, Text Detection, Vulnerability

Detection.

Abstract: The goal of research is to use the Novel YOLO V3 SPP for detecting malicious applications while comparing

it with the OCR technique for computation of access time. Materials and Methods: The Innovative YOLO V3

SPP algorithm is used to determine access time using a sample size of (N=25), a total sample size of (N=50),

and G power is computed to be 80%. In terms of data exploitation prediction, the Novel YOLO V3 SPP has

an access time that is slower (83.36ms) than the OCR algorithm's (79.64ms). According to the results, there

is no statistically significant difference between the Novel YOLO V3 SPP Algorithm and the OCR Algorithm

with p=0.218 (independent sample t-test p<0.05). In comparison to OCR's access time of 79.64ms, the novel

YOLO V3 SPP method predicts vulnerabilities in native programmes with a longer access time of 83.36ms.

1 INTRODUCTION

Deep learning is quickly developing their standards in

text and object recognition all over the world (Zhang

et al. 2022). It's basically working on neural

networking and it's becoming easy to detect all the

solutions to this project. Deep learning basically

works under machine learning in all over internet

access. It is used in most significant manner to find

text in the surrounding images (Garcia-Argibay et al.

2022). It provides a basic ideology of finding

perspective in images. Android applications are

plagued by data exploitation everywhere in internet

access. The model also needed to be in terms of

accuracy (Bergler et al. 2022). As a result, it is used

to collect data more rapidly on a significant scale. To

eradicate data exploitation in native applications,

Data leakage can be detected using deep learning, and

it can be avoided using a variety of techniques.(El-

Amir and Hamdy 2020)(James et al. 2017)

(Sivakumar et al 2022). Speed and security will be

through deep learning in android applications. In this

way, the proposed research aim will be achieved

(PirahanSiah 2019). Data cleansing, computer vision,

customer relationship management, fraud detection,

speech detection, and many other uses are all possible

with deep learning.(Kim 2017).

This proposed work includes numerous research

papers are published in IEEE, 5000 articles from

Springer, 2400 articles from Science Director, and

4090 publications from Google Scholar on text

detection and prediction of data leakage and it have

more citations than 50 times (Amaratunga 2021). The

best cited article is one where the OCR model is a

powerful tool for detecting text in natural

environments and has received significant attention in

the research community, with 77 citations (Ketkar

and Moolayil 2021). One of the key advantages of

this model is its ability to accurately identify text at

various positions and scales, providing insight into

the importance of features for effective text detection.

Overall, the OCR model has proven to be a valuable

resource for researchers studying text detection in

natural situations. The use of deep learning for text

detection in natural environments has significantly

Novel the capabilities of computer processing

applications. This research has paved the way for a

new era of deep learning and has reshaped the way

we approach text detection in natural scenes. Overall,

the advancements in this field have significantly the

effectiveness of text detection using deep learning

techniques (Kelleher 2019). This research focuses on

extracting textual information from natural scenes,

with a particular focus on android applications such

514

T., D. and V., G.

Detection of Malicious Applications Using YOLO V3-Spatial Pyramid Pooling over Optical Character Recognition for Computing Access Time.

DOI: 10.5220/0012571800003739

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

In Proceedings of the 1st International Conference on Artiﬁcial Intelligence for Internet of Things: Accelerating Innovation in Industry and Consumer Electronics (AI4IoT 2023), pages 514-519

ISBN: 978-989-758-661-3

as license plate recognition and navigation. The

findings of this research have been highly influential,

as evidenced by the 26 citations it has received

(Kariya, Fujishima, and Zhang 2019). Overall, the

goal of this research is to improve the ability to

identify and extract relevant text from natural

environments, with the potential to facilitate a range

of practical applications and to provide national

security in internet access (Brum et al. 2023).

Existing systems suffer from certain limitations

such as lesser response time for accessing the data.

(Patterson and Gibson 2017). Depending on the

specific font or script used for the letter, the resulting

matrix is labeled as representing a specific character.

In order to overcome these pitfalls, the proposed

algorithm Novel YOLO V3 SPP is used as an

alternative to OCR algorithm which is basically to

improve the access time. OCR is slower in predicting

the textual materials in images. Also, it does not

provide accurate prediction in small text values

throughout the model (Mueller and Massaron 2019).

It needs vast space to process the model in every

aspect. The quality of image gets lost in processing

the model for national security (Arif Wani et al. 2019)

(G. Ramkuamr et al. 2021). It will not recognize the

printed text only not as hand written text. The goal of

research is to improve the access time using Novel

YOLO V3 SPP comparison with OCR algorithm for

predicting the vulnerability detection and to provide

national security.

2 MATERIALS AND METHODS

The study was carried out in the data analysis lab at

Saveetha School of Engineering, Saveetha Institute of

Medical and Technological Sciences, which has a

high configuration system for carrying out studies and

obtaining results. Two groups were taken into

account for the study, with a sample size of 25.

(Nundy, Kakar, and Bhutta 2021). 80% of the G-

power value, 0.05 alpha, 0.2 beta, and a 95%

confidence range are used to calculate the result

(Kane, Phar, and BCPS n.d.).

CVE (Common Vulnerability and Exposures), a

dataset used in this suggested study, analyses

programmes using predefined threats and

vulnerabilities to generate the vulnerability ratings.

You must use Kaggle to download CVE (Mell 2002).

The dataset used in this analysis helps identify the

vulnerability status. The data set is then subjected to

the methodology devised in this study, and the

outcomes are contrasted with the comparison

algorithm (Brum et al. 2023). Based on the specified

work, these data are logged using real-time apps and

separated into several columns. CVE uses the internet

to enhance national security and is based on a list of

applications that can corrupt data.

In this research, the implementation is carried out

using Jupyter notebook, an open-source web

application that enables the creation and sharing of

documents, including machine learning methods. The

code for this research is saved in the Jupyter

directory, and the system uses darknet-53 to analyze

images in natural scenes. Jupyter notebook allows for

the easy execution and organization of code, making

it a useful tool for researchers and developers. Jupyter

notebook is an open-source web application.

Novel You Only Look Once Algorithm V3 Spatial

Pyramid Pooling Algorithm

Novel YOLO V3 SPP (You Only Look Once V3

Spatial Pyramid Pooling) algorithm is the sample

preparation group 1 and the proposed algorithm for

this proposed work. Novel YOLO v3 SPP is a real-

time object detection algorithm that uses

convolutional neural networks (CNNs) to identify

objects in video. In version 3, the algorithm was

longer to increase both speed and accuracy and to

provide national security. It is designed to detect large

sections of objects at a time, using comparison

methods to quickly identify and classify data as it is

received. The proposed research involves inputting an

image and processing it through CNNs, including 24

convolution and max-pooling layers, followed by

fully connected layers. To reduce the number of

layers, convolution is used followed by layer

convolution. The final layer of Novel YOLO V3 SPP

predicts an output by generating from the fully

connected layers and reshaping them to the size of the

input image in usage of internet access. Overall, the

goal of this research is to improve the speed and

accuracy of object detection, while maintaining a

high level of efficiency to increase national security

in internet access. Table 1 procedure refers to the

procedure of the Novel YOLO V3 SPP algorithm.

Optical Character Recognition Algorithm

Optical character recognition (OCR) is sample

preparation group 2 and is a process and technology

that converts text and images containing text into a

digital format that can be read by a computer. OCR

algorithms work by dividing an image of a text

character into smaller sections and analyzing the

presence or absence of pixels in each section. The

limitations of OCR is that the quality of the image can

be lost and not worth the small amount of text. Also,

it needs lots of space to store an image and even to

process it. Based on the specific font or script used,

Detection of Malicious Applications Using YOLO V3-Spatial Pyramid Pooling over Optical Character Recognition for Computing Access

Time

515

the resulting matrix is labeled as representing a

particular character. This allows OCR to recognize

and transcribe text from images, making it a useful

tool for various applications. OCR also enables the

conversion of large quantities of paper documents

into digital files, improving the efficiency of text

storage, processing, and searching. In summary, OCR

is a valuable tool for digitizing and organizing text-

based information, making it more accessible and

easier to use. The network uses a CNN

(Convolutional Neural Network) approach to classify

the results into two categories, and it is able to

generalize this binary classification effectively. The

future work is to prove that to provide longer access

time, Safe and secure usage will be provided through

this. Table 2 procedure refers to the procedure of

OCR algorithm.

Statistical Analysis

The statistical programme used to calculate the

standard derivation, mean and standard error means,

mean difference, sig, and F value uses IBM SPSS of

version 26.0. The number of installations discovered

in real time is an independent variable, and

application access time is a dependent variable. An

independent T-test analysis is carried out in this

investigation.(Brum et al. 2023).

3 RESULTS

Table 1: Novel YOLO V3 SPP procedure for a novel (You

Only Look Once Version3 spatial pyramid pooling). This

Novel YOLO V3 SPP technique makes it possible to

predict the used data outside of the application quickly. It

also demonstrates the access value much better.

Input: CVE (Common Vulnerability and Exposures) dataset

Output: Longer access time

Step 1: First input is a captured natural scene image.

Step 2: It undergoes through the convolutional layers, it uses 1*1 laye

which predicts the size of the image.

Step 3: Observing the image and calculating the weights in the image.

Step 4: It has three anchors which are called anchors which predicts

the log space.

Step 5: The outcome is computed using access time and picture

detection accuracy.

The Novel YOLO V3 SPP (You Only Look Once)

procedure is shown in Table 1. And information is

collected from the dataset. The novel YOLO V3 SPP

operates by recognising each image in the

surrounding natural scene. These parameters are

specified for additional model testing. must make an

effort to improve the access time. The OCR procedure

is shown in Table 2. The Novel YOLO V3 SPP and

OCR method is compared using the dataset. OCR

predicts less accurately than Novel YOLO V3 SPP. It

is therefore tested using the methods for faster access

time. The raw data table of access time in Table 3

includes both the Novel YOLO V3 SPP and OCR.

Table 2: Procedure for OCR (Optical Character

Recognition). The OCR algorithm takes the dataset of

vulnerability in apps and helps to predict it by using this

algorithm.

Input: CVE (Common Vulnerability Exposure) Dataset

Output: Longer access time

Step 1: Captured image from the natural scene is further divided into

scales.

Step 2: Observes the output and predicts the weight.

Step 3: The image undergoes the neural networks techniques.

Step 4: Dividing images in no of scales.

Step 5: Finally, predicting the output by score generation.

Table 3: Table of raw data for access times Novel YOLO

V3 SPP (You Only Look Once V3-Spatial Pyramid

Pooling) and OCR (Optical Character Recognition).

S.NO

Novel YOLO V3 SPP (You Only

Look Once V3-Spatial Pyramid

Pooling) Access time (ms)

OCR

(Optical Character

Recognition)

(

Access time

)

(

)

1 62 98

2 64 96

3 66 94

4 68 92

5 70 90

6 72 88

7 74 86

8 76 85

9 78 84

10 80 83

11 82 82

12 84 81

13 86 80

14 88 79

15 89 78

16 90 77

17 91 76

18 92 75

19 93 74

20 94 72

21 95 71

22 96 70

23 97 68

24 98 66

25 99 62

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

516

Table 4: Group statistics for contrasting Novel YOLO V3 SPP and OCR Algorithm using independent samples. The mean

access time in Novel YOLO V3 SPP is 83.36 milliseconds, compared to 79.64 milliseconds in OCR.

Algorithm N Mean Std. Deviation Std. Error Mean

Access time

Novel YOLO V3 SPP 25 83.36 11.626 2.325

OCR 25 79.64 9.327 1.865

Table 5: T-Test for Comparing Independent Statistical Samples of Novel YOLO V3 SPP with OCR Algorithm, 95%

confidence interval. It shows that there is no statistical significance difference between the Novel YOLO V3 SPP Algorithm

and OCR Algorithm with p=0.218 (p>0.05).

F Sig t df Sig(2-tailed)

Mean

Difference

Std. Error

Difference

Lower Upper

Access time

Equal Variances

assumed

Equal variances

not assumed

2.743 .104

1.248

45.845

.218

3.720

2.981

2.274

2.281

9.714

9.721

Figure 1: Comparison between Novel YOLO V3 SPP and OCR algorithm based on mean access time. The access time of

Novel YOLO V3 SPP is significantly better than OCR algorithm. X-axis: Novel YOLO V3 SPP Vs OCR algorithm Y-axis:

Mean access time. Error Bar +/- 1 SD.

The values for the Novel YOLO V3 SPP (You

Only Look Once) and the OCR Algorithm are shown

in Table 4 as N (25), Mean (83.36), Standard

Deviation (11.626), and Std. Error Mean values

(2.325) respectively. The T-Test results for

statistically independent samples are shown in Table

5. The Novel YOLO V3 SPP and OCR algorithm has

a mean difference of 3.720, a standard error of 2.981,

and a 95% confidence interval. According to the

results, there is no statistically significant difference

between the Novel YOLO V3 SPP Algorithm and the

OCR Algorithm with p=0.218 (p>0.05). The T-Test

results for the Novel YOLO V3 SPP and OCR

algorithm are shown in a bar graph in Figure 1. It

displays the Novel YOLO V3 SPP and OCR

algorithm's access time. The bar graph gain and

values compute the access time and use it.

4 DISCUSSION

The YOLO V3 SPP access time has a significance

value of 83.36, which is higher than the OCR rate.

The OCR significance value of 73.64 is more than the

significance of the YOLO V3 SPP method in terms of

statistical outcome values. More than (p>0.05) is the

no statistical significance value of 0.103.

The basic principle behind OCR learning is to

combine the predictions of multiple base learners into

Detection of Malicious Applications Using YOLO V3-Spatial Pyramid Pooling over Optical Character Recognition for Computing Access

Time

517

a single result that is expected to outperform any

individual member of the group, along with an

unrelated error on the target data sets (Tejas et al.

2016). Group learning is based on using different

models to solve problems related to accuracy and

access time (Yu 2021). The OCR algorithm is

particularly effective when there is a significant

difference between the models in terms of access time

(Kariya, Fujishima, and Zhang 2019). To avoid errors

due to multiple images being captured at once, the

algorithm removes unwanted data outside the

bounding boxes (Daneshfar, Fathy, and Alaqeband

2018). This also helps to select and predict specific

objects or parts with an accuracy of 86% (Bao et al.

2022). One of the key advantages of OCR is that it is

designed to meet all assumptions consistent with the

training data, with an access time of 85% (Bao et al.

2022). Overall, OCR is a powerful tool for object

detection and prediction, with a range of benefits and

android applications (Xu, Xue, and Zhao 2022).

Factors affecting the research work are

identifying and predicting data leakage in android

applications using various algorithms, including the

OCR algorithm as an object detection tool. The

limitations of OCR is that the quality of the image can

be lost and not worth a small amount of text. Also, it

needs lots of space to store an image and even to

process it. The version 3 of the OCR algorithm can be

used to extract textual data from images, dividing the

data into smaller sections to facilitate quick

processing in android applications. The algorithm

generates a grid for each image in the task and is

useful for predicting text or objects in the virtual

world. The OCR algorithm has many applications,

including traffic control and license plate

identification, as well as speed detection. However, it

has limitations in terms of access time and accuracy,

which may impact future work. Overall, the aim of

this research is to improve the security and usability

of android applications, while providing accurate and

efficient text detection in images. And provide longer

access time. The future work is to prove that to

provide longer access time, Safe and secure usage

will be provided through this in android applications.

5 CONCLUSION

Novel YOLO V3 SPP and the OCR algorithms have

predicted real-time data leakage applications for

Android using various trained detection datasets.

When comparing the two algorithms, Novel YOLO

V3 SPP has a higher access rate than the OCR

algorithm. The Novel YOLO V3 SPP's performance

and sensitivity are superior to the OCR's (83.36ms)

(73.64ms). In comparison to the OCR method, data

loss is reduced in the Novel YOLO V3 SPP.

REFERENCES

Amaratunga, Thimira. (2021). “What Is Deep Learning?”

Deep Learning on Windows.

https://doi.org/10.1007/978-1-4842-6431-7_1.

Arif Wani, M., Farooq Ahmad Bhat, Saduf Afzal, and Asif

Iqbal Khan. 2019. Advances in Deep Learning.

Springer.

Bao, Wuzhida, Cihui Yang, Shiping Wen, Mengjie Zeng,

Jianyong Guo, Jingting Zhong, and Xingmiao Xu.

(2022). “A Novel Adaptive Deskewing Algorithm for

Document Images.” Sensors 22 (20).

https://doi.org/10.3390/s22207944.

Bergler, Christian, Simeon Q. Smeele, Stephen A. Tyndel,

Alexander Barnhill, Sara T. Ortiz, Ammie K. Kalan,

Rachael Xi Cheng, et al. (2022). “ANIMAL-SPOT

Enables Animal-Independent Signal Detection and

Classification Using Deep Learning.” Scientific

Reports 12 (1): 21966.

Brum, Mauro, Luciano Pereira, Rafael Vasconcelos

Ribeiro, Steven Jansen, Paulo R. L. Bittencourt, Rafael

S. Oliveira, and Scott R. Saleska. (2023). “Reconciling

Discrepancies in Measurements of Vulnerability to

Xylem Embolism with the Pneumatic Method: A

Comment on Chen et Al. (2021) ‘Quantifying

Vulnerability to Embolism in Tropical Trees and Lianas

Using Five Methods: Can Discrepancies Be Explained

by Xylem Structural Traits?’: A Comment on Chen et

Al. (2021) ‘Quantifying Vulnerability to Embolism in

Tropical Trees and Lianas Using Five Methods: Can

Discrepancies Be Explained by Xylem Structural

Traits?’” The New Phytologist 237 (2): 374–83.

Daneshfar, F., W. Fathy, and B. Alaqeband. (2018). “A

Metaheuristic Algorithm for OCR Baseline Detection

of Arabic Languages.” Computer Vision.

https://doi.org/10.4018/978-1-5225-5204-8.ch027.

El-Amir, Hisham, and Mahmoud Hamdy. (2020). “Deep

Learning Fundamentals.” Deep Learning Pipeline.

https://doi.org/10.1007/978-1-4842-5349-6_9.

Garcia-Argibay, Miguel, Yanli Zhang-James, Samuele

Cortese, Paul Lichtenstein, Henrik Larsson, and

Stephen V. Faraone. (2022). “Predicting Childhood and

Adolescent Attention-Deficit/hyperactivity Disorder

Onset: A Nationwide Deep Learning Approach.”

Molecular Psychiatry, December.

https://doi.org/10.1038/s41380-022-01918-8.

G. Ramkumar, R. Thandaiah Prabu, Ngangbam Phalguni

Singh, U. Maheswaran, Experimental analysis of brain

tumor detection system using Machine learning

approach, Materials Today: Proceedings, 2021, ISSN

2214-7853,

https://doi.org/10.1016/j.matpr.2021.01.246.

James, J., Lakshmi, S. V., & Pandian, P. K. (2017). A

preliminary investigation on the geotechnical properties

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

518

of blended solid wastes as synthetic fill material.

International Journal of Technology, 8(3), 466-476.

Kane, Sean P., Phar, and BCPS. n.d. “Sample Size

Calculator.” Accessed April 18, 2023.

https://clincalc.com/stats/samplesize.aspx.

Kariya, Ken, Takahiro Fujishima, and Lifeng Zhang.

(2019). “Development of OCR Mobile Application

Including Miss-Recognized Proofreading System

Using Database Search Algorithm.” Proceedings of The

7th International Conference on Intelligent Systems

and Image Processing 2019. https://doi.org/1

0.12792/icisip2019.034.

Kelleher, John D. (2019). Deep Learning. MIT Press.

Ketkar, Nikhil, and Jojo Moolayil. (2021). “Introduction to

Machine Learning and Deep Learning.” Deep Learning

with Python. https://doi.org/10.1007/978-1-4842-5364-

9_1.

Kim, Phil. 2017. “Deep Learning.” MATLAB Deep Learning.

https://doi.org/10.1007/978-1-4842-2845-6_5.

Mell, Peter. 2002. Use of the Common Vulnerabilities and

Exposures (CVE) Vulnerability Naming Scheme:

Recommendations of the National Institute of Standards

and Technology.

Mueller, John Paul, and Luca Massaron. 2019. Deep

Learning For Dummies. John Wiley & Sons.

Patterson, Josh, and Adam Gibson. (2017). Deep Learning:

A Practitioner’s Approach. “O’Reilly Media, Inc.”

PirahanSiah, Farshid. (2019). “Computer Vision, Deep

Learning, Deep Reinforcement Learning.”

https://doi.org/10.14293/s2199-1006.1.sor-

uncat.clzwyuz.v1.

Sivakumar, V. L., Nallanathel, M., Ramalakshmi, M., &

Golla, V. (2022). Optimal route selection for the

transmission of natural gas through pipelines in

Tiruchengode Taluk using GIS–a preliminary study.

Materials Today: Proceedings, 50, 576-581.

Tejas, Bhonsale, Department Of Computer Engineering, A.

I. S. S. M. S. College of Engineering, Pune,

Maharashtra, and India. (2016). “Number Plate

Recognition and Document Verification Using Feature

Extraction OCR Algorithm.” International Journal Of

Engineering and Computer Science. https://doi.org/

10.18535/ijecs/v5i11.84.

Xu, Xing, Zhenpeng Xue, and Yun Zhao. (2022).

“Research on an Algorithm of Express Parcel Sorting

Based on Deeper Learning and Multi-Information

Recognition.” Sensors 22 (17).

https://doi.org/10.3390/s22176705.

Yu, Ling. (2021). “State-of-the-Art OCR Algorithm for

Translation Guiding: An Online-Offline Combinational

Model.” 2021 Third International Conference on

Inventive Research in Computing Applications

(ICIRCA). https://doi.org/10.1109/icirca51532.2021

.9544563.

Zhang, Lisheng, Wentao Wang, Hanqing Yu, Zheng Zhang,

Xianbin Yang, Fengwei Liang, Shen Li, Shichun Yang,

and Xinhua Liu. (2022). “Remaining Useful Life and

State of Health Prediction for Lithium Batteries Based

on Differential Thermal Voltammetry and a Deep

Learning Model.” iScience 25 (12): 105638.

Detection of Malicious Applications Using YOLO V3-Spatial Pyramid Pooling over Optical Character Recognition for Computing Access

Time

519