Detecting Fake Banknotes: Performance Evaluation of ML and DL

Algorithm

Gujarathi Kalyani

, Basinepalli Keerthi

, G. Shaheen Firdous

Boya Vasavi

and Malipeddi Likhitha

Department of CSE(AI), Ravindra College of Engineering for Women, Kurnool, Andhra Pradesh 518002, India

Department of CSE, Ravindra College of Engineering for Women, Kurnool, Andhra Pradesh 518002, India

Keywords: Counterfeit Detection, Machine Learning, Deep Learning, Convolutional Neural Networks (CNN), Financial

Security.

Abstract: For financial security, making sure counterfeit banknotes are detected is important. We evaluate the

performance of many Machine Learning (ML) and Deep Learning (DL) algorithms to deceive fake currency

accurately. It proposes extracting the numerical and visual features variance, skewness, entropy, of the

wavelet transformed images which are fed to train and test the classification models. Important algorithms,

such as Support Vector Machines (SVM), Decision Trees, Random Forests and Neural Networks are

implemented and compared with respect to performance metrics like accuracy, precision, recall and F1 –

score. Also, the detection accuracy is improved by using deep learning models, i.e., Convolutional Neural

Networks (CNNs), which are capable of automated feature extraction. For the analysis, the dataset is used

which contains labeled instances of genuine and counterfeit banknotes. Strengths and limitations of each

approach are discussed and the applicability to the real word is discussed. Accuracy and robustness in

counterfeit note detection using dummy models of Random Forest and deep learning models, e.g. CNNs, are

superior according to results. The potential of AI driven solutions in automating counterfeit detection has been

established in this project as it is a scalable, efficient, and cost-effective solution for the banking industry. The

advancement of secure and reliable financial systems is made by leveraging data driven

technologies in this study.

1 INTRODUCTION

In fact, counterfeiting remains a threat to the global

economy as it undermines the financial system and

brings about massive losses (Zhang & Huang, 2018).

Counterfeiters are employing advanced techniques

making traditional ways to detect counterfeit

banknotes (such as visual inspection and utilizing

physical security features such as water marks,

security threads and ultraviolet markings) less

reliable in identifying potential counterfeit banknotes

(Li & Liu, 2017). All these processes are human

dependent processes and hence susceptible to errors,

slow in execution and inconsistent making them

impossible for high volume spaces such as banks,

automated teller machines (ATMs), store etc. (Sahoo

& Behera, 2020). We need to develop innovative

solutions to protect financial transactions and keep

the public trust in currency in view of the increasing

scale and sophistication of the counterfeiting on our

currency (Tian & Li, 2020).

The problems stated above are addressed in this

research through applying machine learning (ML)

and deep learning (DL) algorithms to establish

automated, accurate and efficient counterfeit

banknote detection systems (Yin & Li, 2020). Unlike

conventional approaches which rely on predefined

features, ML and DL techniques are capable of

analysing sophisticated patterns existing in banknote

images and, therefore, pointing out elusive features

that differentiate between true and forged notes

(Akkus & Xu, 2019). This study attempts to identify

most appropriate algorithms (e.g. decision trees,

random forests, convolutional neural networks) and

features (e.g. texture and color variations) in real time

authentication (Borges & Silva, 2021). The long-term

goal is to create deployable robust systems, whether

it be in ATM software, at the bank verification

terminals or at the retail checkout systems to perform

468

Kalyani, G., Keerthi, B., Firdous, G. S., Vasavi, B. and Likhitha, M.

Detecting Fake Banknotes: Performance Evaluation of ML and DL Algorithm.

DOI: 10.5220/0013900100004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

468-475

ISBN: 978-989-758-777-1

the instant checks without having human supervision

involved. Unlike existing approaches, this work

presents the first systematic comparison of several

ML and DL models to find the best solution and

provides a practical framework of improving the

capability of counterfeit detection beyond current

level. These systems offer the promise that they will

help reduce the occurrence of financial fraud, beef up

their security, and restore confidence in monetary

systems all around the globe, which is definitely

something that is sorely needed in today’s technology

filed financial landscape (Yin & Li, 2020; Borges &

Silva, 2021).

2 RESEARCH AREA

2.1 Data Collection and Preprocessing

To start, a dataset of images of banknotes that are real

and fake is obtained. To this end, publicly available

databases or datasets particular to this case will

comprise high resolution images of different

denominations of banknotes from different countries

(Yadav & Verma, 2018). The format of the images is

made standard along with its size and resolution. To

increase the variability of data, some coupled image

augmentation techniques, e.g., rotation, scale change,

and noise addition, are performed to prevent

overfitting (Chen & Liu, 2017). Therefore,

transformations like the grayscale conversion and the

histogram equalization can be applied to improve

contrast of the banknote images and simplify

identification of texture, edges and the fine features

and details (Sharma & Kumar, 2019).

2.2 Feature Extraction

After preprocessing of the images, in the process of

ML based counterfeit detection the subsequent step

after the preprocessing of the images is feature

extraction which plays a very crucial role. The

pertinent features which can be edges, texture, and

color patterns are to be manually extracted in legacy

machine learning type of models like SVM, KNN,

and Random Forests using methods like HOG

(Histogram of Oriented Gradients), Gabor Filters, and

SIFT (Scale-Invariant Feature Transform) (Arora &

Sharma, 2018). In deep learning models types like

CNNs, feature extraction is not required through the

convolutional layers of the network, which learn

hierarchical features from the raw image data (Tan &

Duan, 2021).

2.3 Model Development and Training

At this phase, we implement multiple ML and DL

algorithms to train the models in order to identify

counterfeits. We use manually extracted features to

train the SVM model and kernel functions such as

linear or radial basis function (RBF) to provide better

classification (Chen & Liu, 2017). KNN and Random

Forest are also trained from the extracted features,

where KNN predicts based on closeness to nearest

neighbours and Random Forest generates an

ensemble of decision trees for hard classification

(Sami & Gaurav, 2019). The CNN model, on the other

hand, comprises several convolutional layers to learn

and extract features automatically and dense layers

for the final classification into real or fake classes

(Vijay & Kaur, 2019). The models are trained on a

training data set and cross validation methods used to

validate them in an effort to make them generalizable

(Zhou & Wang, 2016).

2.4 Model Evaluation

After training, the models' performances are

evaluated through a series of metrics: accuracy,

precision, recall, F1-score, and confusion matrix (Tan

& Duan, 2021). These enable one to see how well

each algorithm detects fake currency and

distinguishes it from real notes. The evaluation also

includes testing the models on an independent test set

that was not used in training. The CNN model, being

a deep learning-based approach, ought to perform

better than the standard ML models on accuracy in

terms of its ability to learn complex patterns from

images automatically (Vijay & Kaur, 2019).

However, all models are compared to determine the

most computationally efficient algorithm in terms of

computational resources, training time, and

classification performance (Chen & Liu, 2017).

Further, how different preprocessing steps, e.g.,

image resizing or color correction, influence the

pipeline is examined to determine the optimal

pipeline for counterfeit detection (Sharma & Kumar,

2019). This approach allows for an end-to-end

evaluation of the performance of various ML and DL

techniques, yielding valuable insights into the

practical applicability of the technologies in fake

banknote detection (Arora & Sharma, 2018).

3 EXISTING SYSTEM

Most current methods of counterfeit banknote

detection rely on physical examination and security

Detecting Fake Banknotes: Performance Evaluation of ML and DL Algorithm

469

features such as watermarks, UV stamps, security

threads, and holograms. These outdated methods have

been in practice for decades and are still applied by

the majority of banks, retail stores, and ATMs today.

However, they possess enormous limitations when it

comes to identifying advanced counterfeit money,

which can replicate or replicate security features

3.1 Manual Inspection

The most primitive mode of counterfeit identification

is manual checking, where people are looking at

physical characteristics of banknote to verify his

authenticity. That requires checking the texture, color,

and the security elements like holograms, watermarks

or raised ink.

3.2 Machine-Based Detection

Machine technology in the guise of currency and UV

light detectors are employed to detect counterfeit

notes by detecting visible security features when held

under ultraviolet light. Such machines usually check

for presence of UV marks, embedded filaments, or

invisible watermarks. Even though such machines are

quicker and more precise than their human

counterparts, they cannot detect sophisticated

counterfeit notes that imitate such features.

Counterfeiting has become sophisticated and forgers

are able to now replicate the UV-sensitive features so

that such machines are not helpful in certain

situations. Optical Character Recognition (OCR)

Optical Character Recognition (OCR) is employed to

scan the serial numbers, words, and markings on

banknotes. Such a method enables a quick check

against a database of authentic serial numbers, but it

checks only whether a particular bill is authentic or

not based on the information present in the image.

Sophisticated counterfeit notes, however, may

precisely manufactured serial numbers and words, so

OCR-based systems are poor at identifying

counterfeit bills with very similar look to real bills.

3.3 Feature-Based Machine Learning

Algorithms

Machine learning (ML) is utilized by some systems

algorithms to identify counterfeit banknotes by

examinations on certain features like textures, edges,

and color patterns. Support Vector Machines (SVM),

K-Nearest Neighbors (KNN), and Random Forest

(RF) were utilized to feature-based classification in

counterfeiting. These systems extract certain features

from images and use to recognize a banknote as real

or counterfeit. Nevertheless, such systems remain

susceptible to hand feature extraction and are less

effective in identifying very small patterns that can

differentiate genuine and fake notes, particularly with

the development of counterfeiting methods.

3.4 Limitations of Existing Systems:

Although the existing systems are good, they have

some limitations:

3.4.1 Only Effective Against Basic

Counterfeiting Methods: Traditional systems work

mostly against rudimentary counterfeits that do not

try to replicate advanced security features.

Sophisticated counterfeit notes with carefully

replicated security features can easily bypass most of

these systems

3.4.2 Human Judgment Based

Human inspection is highly reliant on human

experience, which is unreliable, time-consuming, and

prone to errors, especially in high-pressure

environments like banks and shopping malls.

3.4.3 Not Effective for Large Volumes

Manual and mechanical methods are tedious in

dealing with large sums of money, i.e., in ATM

machines, department stores or in money processing

4 PROPOSED SYSTEM

Overview.

The system suggested for detecting counterfeiting

banknotes utilizes the most recent Machine Learning

(ML) and Deep Learning (DL) techniques to enhance

accuracy, efficiency, and scalability in currency

authentication. The objective is to autonomously

detect counterfeit notes in real-time without the

limitations of existing practices, e.g., reliance on

human verification, vulnerability to environmental

conditions, and susceptibility to sophisticated

counterfeiting techniques. The suggested system

involves multiple stages of data acquisition, feature

extraction, training of models, and deployment, with

a focus on leveraging Convolutional Neural

Networks (CNNs) in perdurable image-based forgery

detection.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

470

4.1 Data Collection

The data collection process for detecting fake

banknotes involves gathering a diverse dataset of

images featuring both genuine and counterfeit

banknotes. These images are captured under various

lighting conditions, angles, and resolutions to ensure

variability in the dataset. Additionally, the system

collects visual features from the banknotes, such as

texture, color patterns, security elements like

holograms, watermarks, and serial numbers, which

help in distinguishing authentic from fake notes.

Metadata such as the denomination, country of origin,

series, and year of issue is also extracted to further

assist in identification. Furthermore, each banknote

image is labelled as either genuine or counterfeit,

providing essential annotations for supervised

learning and model training.

4.2 Preprocessing and Feature

Extraction

Different preprocessing techniques are applied to

collected banknote images in order to enhance and

highlight to features for analysis. Grayscale

conversion, image normalization and edge detection

methods are used to focus attention to watermarks,

microtext, or other security features that are encoded

in the banknotes. Convolutional Neural Networks

(CNNs) learn relevant patterns directly from the raw

images automatically in their task of feature

extraction and do not require any manual intervention

for learning relevant features. It may learn the

features such as texture patterns, the quality of prints,

ink distribution, holograms and any other subtle

features that can help us to differentiate the genuine

from counterfeit notes.

4.3 Deep Learning Model (CNN) for

Detection

The method presented here for the purpose of fake

banknote detection is a proposed automatic system

whose approach is to utilize Convolutional Neural

Networks (CNNs) to automatically learn and extract

main features from a large database of images

labelled as real or counterfeit currency. This task is

very appropriate for CNNs to perform, as they are

capable of handling complex patters as well as

hierarchical.

It does not rely on manual extraction of features

from raw images. The convoluted architecture

includes layers which identify convex patterns like

edges or textures, pooling layers to reduce the

dimensions and keep important information and the

fully connected layers for the final classification. This

setup enables the system to process images at great

speed and make accurate decision w.r.t whether the

banknote is real or fake. The top advantage of CNN

is its ability to capture fine and explicit features like

printing in cohesion, ink distribution, watermark and

hologram that is necessary for detecting counterfeit

banknotes. However, capturing these features is

difficult using traditional methods and combined with

the CNNs they are therefore suitable for encoding

task. In addition, the system generalizes well when

the training is done on diverse datasets and the new

unseen banknotes can even include advanced

counterfeiting techniques. High accuracy and

adaptability to real world application are thus ensured

and the system can detect the counterfeit currency for

different currency designs and also for different

counterfeit methods.

4.4 Hybrid Approach for Enhanced

Accuracy

The Proposed System can also be enhanced by adding

various Ensemble Methods such as Random Forest

(RF) or Support Vector Machine (SVM), as a means

of classification refinement. The traditional machine

learning techniques are useful in handling corner

cases or types of counterfeits for which CNN fails

such as print quality variance or counterfeits with

intricate patterns. Such a system has the advantage of

combining the strengths of both deep learning and

traditional machine learning approaches, which

results in a more robust system and allows it to detect

more counterfeit banknotes in a wider variety of

counterfeit banknotes, thus improving the overall

performance and reliability in the real world.

4.5 Model Testing and Tuning

The effectiveness of the system in classifying genuine

versus counterfeit banknotes is going to be evaluated

by a systematic set of metrics, including accuracy,

precision, recall, and F1-score. Furthermore, the

model will be tested for robustness by displaying it

different counterfeit notes of various qualities and

printing techniques to assess its performance under

various conditions. Hyperparameter tuning along

with the use of methods like cross validation will be

performed to optimize the model so it can be fine-

tuned for the best performance under all conditions

and this will make the model accurate in identifying

the counterfeit bills in real scenarios.

Detecting Fake Banknotes: Performance Evaluation of ML and DL Algorithm

471

4.6 Real-Time Deployment

The system is proposed for real time use on

environments such as retail stores, ATMs as well as

banks where counterfeit detection is imperative. This

model will be deployed to process cloud or edge

devices that will be capable of processing image of

the banknote in real time to give them an immediate

fidelity (either a genuine or counterfeit note). Existing

financial infrastructure, like the ATMs, cash counting

machines or self-service kiosks can be easily

integrated with the checks performed to ensure

authenticity of bank notes will be seamlessly

automated with this. The system will further have a

web interface or a mobile application for monitoring,

management, and operational oversight to ensure that

the deployment and maintenance of the system can be

done efficiently in different settings. The system

architecture is shown in figure 1.

4.7 Benefits of the Proposed System

4.7.1 Real Time Detection

The system is intended for real time deployment

application such as retail stores, ATMs, banks etc in

order to detect counterfeits notes with instantaneous

reaction. Herein, it enables users to get feedback

immediately in regards to the genuineness of a

banknote under their scrutiny, therefore, reducing the

flow of dubious banknotes on the market; as well as

improving the speed with which a cash handling

operation is conducted.

4.7.2 Scalability and Integration

This system may be integrated into the existing

financial infrastructure, like ATM, cash counting

machine or self-service kiosk to automatically

process bank note. Its scalability ensures its ability to

be deployed in a setup that ranges from a small

business to several large financial institutions since it

helps to increase efficiency in operation across

different sectors.

4.7.3 Increased Financial Security

For common consumers, the system decreases the

likelihood of getting fake currencies in exchanges and

makes it unlawful for them, lest they waste their own

financial belongings. By continuous verification of

bank notes only genuine currency is in the circulation

and it keeps improving financial security at the

individual level.

5 SYSTEM ARCHITECTURE

Figure 1: System architecture.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

472

6 CONCLUSIONS

Detecting forged banknotes is a crucial challenge for

banks, retailers, and economies around the world.

Traditional methods, such as manual checks and

relying on physical security features, have long been

the go-to solutions for identifying counterfeit

currency. However, as counterfeiters become more

sophisticated and develop advanced techniques to

replicate security features, these traditional methods

are increasingly ineffective. This is where machine

learning (ML) and deep learning (DL) technologies

offer significant promise. These advanced

technologies provide highly accurate, automated

solutions that can detect counterfeit banknotes with

impressive speed, efficiency, and reliability.

In this research, several machine learning and

deep learning techniques such as Support Vector

Machines (SVM), KNearest Neighbors (KNN),

Random Forest (RF), and Convolutional Neural

Networks (CNN) were explored for counterfeit

detection. The comparison of these methods clearly

demonstrated the superior performance of deep

learning, particularly CNNs, in handling complex

image patterns and achieving higher accuracy. CNNs

excel automatically extracting high-level features

from raw images of banknotes, which makes them

particularly well suited for distinguishing between

genuine and counterfeit currency.

By implementing these advanced algorithms,

counterfeit detection systems can be made more

efficient and reliable for real-world applications,

including ATMs, banks, and shopping malls. These

systems can eliminate the need for human

intervention, reducing the risk of human error, and

provide nearly real-time authentication of currency,

thereby enhancing security and minimizing financial

fraud. Overall, the use of ML and DL in anti-

counterfeiting efforts represents a significant step

forward compared to traditional methods. While

further refinement of these models may be required

for specific applications, the potential to improve

accuracy and detection capabilities is immense. As

these technologies continue to evolve, they will make

the detection of counterfeit banknotes faster, more

efficient, and accessible, helping to safeguard

financial systems and economies worldwide

7 RESULT

Figure 2 shows the interface of user to upload the

image and Figure 3 shows the browsing image.

Figure 2: Initially no image is displayed and user is asked to insert image.

Figure 3: Browsing image.

Detecting Fake Banknotes: Performance Evaluation of ML and DL Algorithm

473

Figure 4 shows the input from the user and Figure 5

shows the processing image.

Figure 6 shows the output of real note and Figure 7

shows the fake note.

Figure 4: Input Image of Currency Note.

Figure 5: Image Sent for Processing...

Figure 6: GUI Showing final result (Real note).

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

474

Figure 7: GUI Showing final result (Fake note).

REFERENCES

Akkus, Z., & Xu, Y. (2019). Detecting counterfeit currency

using deep learning. Journal of Machine Learning in

Financial Systems, 4(3), 110-118.

Arora, P., & Sharma, R. (2018). A survey on AI-driven

banknote verification systems. Proceedings of the 2018

International Conference on Artificial Intelligence and

Data Engineering, 102-107.

Borges, R. M., & Silva, P. A. (2021). An improved

approach to fake banknote detection using

convolutional neural networks. International Journal of

Image Processing, 15(1), 52-64.

Chen, X., & Liu, Y. (2017). A comparative study of

machine learning algorithms for fake currency

detection. Journal of Computer Science, 42(6), 156-

162.

Li, J., & Liu, F. (2017). Pattern recognition methods for

banknote verification. Pattern Recognition Letters, 105,

12 18.

Sahoo, A., & Behera, S. (2020). A hybrid technique for

counterfeit currency detection. International Journal of

Electronics and Electrical Engineering, 8(4), 27-34.

Sami, W., & Gaurav, A. (2019). A novel approach to

identifying counterfeit banknotes using AI.

International Journal of Computer Vision and Pattern

Recognition, 6(4), 45-52.

Sharma, V., & Kumar, A. (2019). Image-based

authentication of banknotes using neural networks.

Journal of Artificial Intelligence and Image Processing,

9(5), 281-295.

Tan, J., & Duan, X. (2021). Deep learning-based counterfeit

banknote classification. Expert Systems with

Applications, 175, 114832.

Tian, Y., & Li, M. (2020). Image-based banknote

authentication using machine learning. Journal of

Computer Science and Technology, 35(3), 412-423.

Vijay, P., & Kaur, G. (2019). An automated system for fake

currency detection using deep learning. IEEE Access,

7, 63108-63117.

Yadav, A., & Verma, R. (2018). Counterfeit banknote

detection using image processing techniques.

International Journal of Computer Science and

Information Security, 16(12), 113-118.

Yin, Y., & Li, Y. (2020). Fake currency detection using

machine learning algorithms. International Journal of

Advanced Computer Science and Applications, 11(4),

160 167.

Zhang, J., & Huang, Y. (2018). A study on financial fraud

detection techniques. Journal of Financial Technology,

6(2), 91-98.

Zhou, W., & Wang, C. (2016). Real-time counterfeit

detection in banking transactions. Proceedings of the

2016 IEEE International Conference on Information

and Automation, 1347-1351.

Detecting Fake Banknotes: Performance Evaluation of ML and DL Algorithm

475