2023 International Conference on Communication, Circuits, and Systems (IC3S), May 26-28, 2023, Bhubaneswar,

India

An Explainable Machine Learning Model to

Analyze and Detect Fake Currency
Parnika J Snigdha Sen Divya R
Dept. CSE, GAT Dept. CSE, GAT Dept. CSE, GAT
Global Academy Of Technology Global Academy Of Technology Global Academy Of Technology
2023 International Conference on Communication, Circuits, and Systems (IC3S) | 979-8-3503-2590-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/IC3S57698.2023.10169455

Bengaluru, India Bengaluru, India Bengaluru, India

[email protected] [email protected] [email protected]

Deepika S Dharshini V
Dept. CSE, GAT Dept. CSE, GAT
Global Academy Of Technology Global Academy Of Technology
Bengaluru, India Bengaluru, India
[email protected] [email protected]

Abstract— The advent of state-of-the-art techniques and approach to aid in the development of applications that
technologies has exposed human beings to some threats as well, facilitate currency detection. Using Machine Learning(ML)
apart from providing several advantages. Fake currency is such algorithms we can build models to detect fake notes without
a major threat to society which in turn obstructs the country’s manually analyzing them and hence it will be reducing human
economic growth and financial status and causes an impact on effort. The model learns from the dataset fed and assists to
human life. In addition to it, detecting fake currency is very detect a counterfeit note after being trained.
challenging for the human eye. Detection of such currency is an
important task and also consumes a lot of time if done manually. Since fake currency is detrimental and has a
To address this critical issue, our paper contributes to degrading effect on the society, economic and financial status
identifying fake currency using Machine learning (ML) of the country, detecting these counterfeit notes is as important
algorithms. In our article, we experimented with multiple ML as the value of the real currency. Though fake currency
algorithms to detect fake currencies and analyze their detection machines are available, it is not affordable for
performance. Among the six algorithms experimented with for common people. In this paper, to overcome the challenging
our work, Random Forest has shown the highest accuracy with task of fake currency detection manually or with expensive
99.27% and a precision of 0.99 compared to the other machines, we use Machine Learning(ML)-based approaches.
algorithms. Additionally, we have demonstrated the The experiments performed in this paper involve data
interpretability of the ML model using LIME and SHAP which
containing the features of both real and fake images. These
would be beneficial in understanding the important factors
features help us in classifying the currencies well. The
behind the particular prediction. We also performed
hyperparameter tuning using GridSearchCV and manually too
algorithms considered are Logistic Regression (LR), Random
which showed us an improved accuracy performance of 99.99% Forest(RF), Decision Trees(DT), Neural Networks(NN),
from 99.27% for Random Forest. Feature Importance analysis Support Vector Classifier(SVC), and Extra Trees
and Explainability concepts are also demonstrated to provide an Classifier(ETC). These algorithms can easily handle numeric
in-depth analysis of model prediction. features. Hence we have done a comparative study of their
performance and chosen the best model to classify the
Keywords—Machine Learning, GridSearchCV, Random currencies.
Forest Classifier, Hyperparameter Tuning, Explainable AI
The remainder of the section is organized as follows:
I. INTRODUCTION Section II describes the Background Study, the Methodology
used for our study is described in Section III, the Results and
In today’s world, Machine Learning Algorithms are
Discussion are described in Section IV, Section V describes
being used in various domains such as retail, health care [1],
the Explainable AI-based Machine Learning(ML) model, and
astronomy[2], medicine, agriculture, speech recognition,
our study is concluded in Section VI as Conclusion.
email filtering, traffic prediction, self-driving automobiles,
and many more. With today's advanced printing and scanning II. BACKGROUND STUDY
technology, anyone can produce counterfeit currency. These
notes are produced without the legal sanction of states or Multiple approaches have been already proposed in
governments. The manufacturing of fake currency can harm a this area. Some of them are discussed in this section.
nation's economy. The value of real money is diminished by Vivek Sharan et al. in their paper[3] discussed four
fake money, which has an impact on consumers through supervised machine learning algorithms—SVC, K-Nearest
inflation, leads to monetary losses for companies, and also Neighbors (KNN), DT, and LR along with image processing
fuels the spread of terrorism. Also, it is difficult for ordinary methods to categorize Indian currency notes as authentic or
people to identify fake currency. Therefore fake currency false. These methods were applied to the dataset from the UCI
detection has become a very challenging and crucial task. In machine learning repository. Additionally, the effectiveness
recent times, machine learning and deep learning techniques of each algorithm is evaluated in terms of precision, accuracy,
have shown tremendous performance in fake currency F1-score, and recall. KNN algorithm achieved superior results
detection. These techniques require minimum human to other algorithms.
interference and make the job of humans easier. Through
automated systems and algorithms and machine learning

979-8-3503-2590-4/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 09,2025 at 11:04:26 UTC from IEEE Xplore. Restrictions apply.
 Pravidhi Garkoti et al. in their paper[4] discussed image approaches worked. For each method, the confusion matrix is
processing methods to suggest an automated monetary system. produced using the cross-validation predict module, and data
Pre-processing techniques are carried out, such as blurring and were plotted to visualize the work of these algorithms. The
noise reduction, when cash is scanned to produce an image. model is assessed using performance indicators including
The feature extraction procedure is the last one before the last accuracy, precision, and f-score.
step, which is the presentation of the result, which is finished M. Jadhav et al. in their paper[14] used deep learning
after border detection and cropping. These steps facilitate the to distinguish between fake and real currency notes. They
processing of money recognition and simplify the process of employed the MATLAB program to do automatic detection of
identification for humans. fraudulent and authentic Indian and foreign currency using a
P. Ponishjino et al. in their paper[5] proposes utilizing deep learning approach.
image processing to identify counterfeit cash. They will use
two input pictures in the system. During Image Processing, the III. METHODOLOGY
image was modified and filtered using a medium filter. The In our proposed system, the data is collected from
image was then transformed from RGB to HSV to get the [15] and its description is given in the following section. The
contrast value. Following conversion, the edges of real and data is visualized using Seaborn Heatmap to check the missing
fake currencies are discovered. values. Since our data set doesn’t have any missing values, it
A. H. Ballado et al. in their paper[6] validate the is then split for training and testing purposes in the ratio of
application of canny edge technology for picture enhancement 80:20. This implies 80% of the data is utilized to train the
and image sensing. Optically Variable Device (OVD) was model, and 20% is used to test it. The data collected is
successful in detecting counterfeit Philippine banknotes. The standardized using a StandardScaler. Using the classification
image was captured and processed using canny-edge algorithms and the data set, the model is built. 6 algorithms
technology by MATLAB GUI software. are considered for our study. Then the performance metrics
Ahmed et al. in their paper[7] introduced a basic are calculated to obtain the best algorithm. The best algorithm
software that can detect fraudulent currency from photos. It is evaluated and tuned for the best results and also used for
has various limits in terms of technology and the printing building the explainable model. LIME and SHAP are used to
standards of Bangladeshi paper currency, but it can pose a depict the explainability of the model. The algorithm is built
significant danger to cash counterfeiting. and tested on Google Colab. Google Colab is an online cloud
S. Murthy et al. in their paper[8] discussed a system platform that allows us to execute the Python code and analyze
that employs image processing techniques. The suggested the data for Machine Learning. Fig. 1 depicts the workflow of
methods for classifying denominations and identifying our proposed system.
counterfeit notes are 100% and 90% accurate, respectively.
Alekhya et al. in their paper[9] intends to search for
and extract visible and hidden signs on paper cash for effective
categorization. They employed the MATLAB approach. It
will assist the ordinary man in controlling the circulation of
counterfeit cash in our nation by using a MATLAB approach
in a mobile phone with a scanner or a camera to identify
counterfeit notes.
Tushar Agasti et al. in their paper[10] discussed a
basic approach that worked properly. The image processing-
based fake currency detection system was built in MATLAB.
The method begins with picture acquisition and ends with the
computation of the intensity of each extracted characteristic.
It is considered original money if the computed intensity is
more than 70%.
Colaco et al. in their paper [11] describe the strategy that
was used to implement the machine that identifies false notes
using image processing. They identify the validity of money
notes using OpenCV software, as well as several hardware
components, including an 8MP Raspberry Pi camera, a
Raspberry Pi 3 B+, and a box setup. The characteristics are
compared depending on their intensities (pixel count). The
number of matching features determines whether the currency
is real or counterfeit.
S. R. Darade et al. in their paper[12], an automated
method is built to identify Indian currencies and determine
whether they are fake or genuine. The image processing
approach is used to identify fake Indian rupee notes. The
technology generates accurate, reliable results. The input is Fig. 1. Flowchart of Proposed System
recorded by a CCD camera in this system, and the output is
displayed on a PC. A. Data Set
A.Bhatia et al. in their paper [13] presents a method for A data set is information regarding a particular topic
detecting counterfeit cash using KNN and image processing. that can either be organized or unorganized. The initial data
They also used SVC, and Gradient Boosting Classifier to train set was created by [15] on capturing the images of both
the models, and they compared how well these three original and counterfeit banknotes and then compressing those

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 09,2025 at 11:04:26 UTC from IEEE Xplore. Restrictions apply.
images using a wavelet transform. The data set contains four 5) SVC (Support Vector Classifier):
key features from each compressed input image and a feature The purpose of a Linear SVC is to categorize the data
that categorizes each currency as either genuine or fake. The by generating a "best fit" hyperplane. After getting the
size of the dataset is 44 KB. As depicted in Fig. 2, the dataset hyperplane, then feed some attributes to the classifier to
that we have taken consists of 1372 rows and 5 columns along acquire the "predicted" class.
with the column headers. The dataset consists of 5 attributes 6) Extra Tree classifier:
which are variance, skewness, curtosis, entropy, and class. Extra tree classifiers aggregate the classification
Variance, skewness, curtosis, and entropy are continuous outcomes of several de-correlated decision trees grouped in a
attributes in this dataset that represent the variance, skewness, "forest" to produce a single classification outcome. They are a
curtosis, and entropy of wavelet-transformed images type of group learning technique. This method produces a
respectively. Class is a categorical attribute. If the class is 0 large number of unpruned decision trees from the training
then the currency is real, otherwise fake. Based on the input dataset. Faster and more accurate than random forest.
attributes, a particular attribute called a label (this can also be
called a class label, output, prediction, target, or response) will C. Our Contribution
be predicted. In this dataset class is the target variable. The The data set collected consists of a target attribute class
dataset has no missing values. which is a categorical value. Because of this, we used
classifiers that were imported from the scikit-learn library and
calculated the performance metrics like accuracy, precision,
training time, prediction time, F1 score, and recall of each of
the above-mentioned algorithms.
The algorithms are run with the data as input and the
values are noted down as shown in Table I. The following
formulas are used for calculating the above-mentioned
Fig. 2. Sample Dataset performance evaluation metrics:
• Accuracy=(TP+TN)/(TP+TN+FP+FN)
B. Algorithms • Precision=TP/(TP+FP)
Supervised machine learning algorithms are a set of • Recall=TP/(TP+FN)
functions that help us to learn patterns of data from given input • F1 score=2((Precision*Recall)/(Precision + Recall))
and output through functional mapping. These are the six where TP=True Positive, TN=True Negative, FP=False
algorithms that we have considered because of their popularity Positive, FN=False Negative, and F is the F1 score.
and their major use in solving classification problems:
TABLE I. PERFORMANCE METRICS CALCULATED FOR ALGORITHMS
1) Logistic Regression: Algorithm Accur Prec Traini Predic F1 recall
A machine learning method called logistic regression acy ision ng tion scor
is used to determine how likely it is that a given instance (In Time Time e
belongs to a certain class. If the probability exceeds 50%, the percen (In (in
model returns a positive case; otherwise, it predicts a negative tage) second second
s) s)
instance. Therefore, it is labeled as a binary classifier. Logistic 98.54 0.98 0.008 0.002 0.99 0.99
2) Random Forest Classifier: Regression
A supervised machine learning technique known as Random 99.27 0.99 0.16 0.012 0.99 0.99
a "random forest classifier" includes a collection of decision Forest
trees selected at random from a subset of the training set, Classifier
Decision 98.90 0.99 0.011 0.004 0.99 0.99
which then adds up the votes from the different decision trees Tree
to provide the final prediction. The accuracy increases and Neural 97.45 0.97 3.829 23.189 0.97 0.98
with increasing forest size, the likelihood of overfitting Network
reduces. SVC 98.54 0.98 0.028 0.009 0.99 0.99
3) Decision Tree: Extra Tree 83.27 0.87 0.12 0.022 0.83 0.80
A supervised machine learning model that is simple Classifier
to grasp is a decision tree. Both classification and regression
trees can be done by it. As implied by the name, a decision As the Table I depicts, the Random Forest Classifier has
tree is created using a binary tree data structure. It is an the highest accuracy when compared with the other algorithms
approach for approximating distinct-valued target functions in considered. Logistic Regression is the algorithm that is trained
which a decision tree serves as a representation of the learning the fastest and has the least prediction time. Fig. 3 depicts the
function. accuracy comparison of the algorithms.
4) Neural Network:
One form of artificial intelligence called a neural
network educates computers to comprehend information in a
manner similar to that of the human brain. This approach may
be used to uncover hidden relationships and patterns. An
abstraction might be a depiction of the input data itself in the
hidden levels. These layers help the neural network
understand various parts of the content by employing their
own internal logic.

Fig. 3. Accuracy Comparison of Algorithms

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 09,2025 at 11:04:26 UTC from IEEE Xplore. Restrictions apply.
 Since the Random Forest Classifier has the highest values that achieves better performance on the data set. There
accuracy we have considered the Random Forest Classifier for are different methods to find this combination like
further analysis. We have performed feature importance to RandomizedSearch and Bayesian Optimization. But since
identify which feature contributed the most to the prediction. GridSearchCV provides output effectively, we have
Further to increase the accuracy of the algorithm, experimented with it.
Hyperparameter tuning is carried out both manually and using
GridSearchCV. An additional metric, ROC Curve, is plotted. Hyperparameters and the values obtained using
To interpret how our algorithm came to a result and whether GridSearchCV are :
• max_depth -> It represents the maximum level of each tree
the result obtained can be trusted or not, we have implemented
Explainable AI which includes LIME and SHAP. so that it performs well and captures most of the
information from the training data. (max_depth = None)
D. Feature Importance • max_features -> For each splitting, this parameter
Since the Random Forest Classifier has the highest represents the maximum level of features
accuracy, we consider this algorithm to perform further tried.(max_features = 1)
analysis. The feature importance graph for Random Forest • n-estimators -> It represents the number of decision trees
Classifier algorithm is plotted to determine the attribute with to be used. (n_estimators = 100)
the highest Feature score. A feature importance graph is • min-samples-leaf ->At the leaf node of each tree, this
plotted to know the impact of each of the features on the model parameter indicates the minimal amount of samples
contributing to the outcome. The Feature Importance of needed. (min_samples_leaf = 1)
Random Forest is plotted as shown in Fig. 4. This method is • min-samples-split -> To split the internal node of each tree,
also an approach to Explainable AI, which is covered in this parameter indicates the minimal amount of samples
further sections. Here, the attribute variance has the highest needed. (min_samples_split = 3)
feature score with a value of 0.573198. This means that
variance has the highest impact on the model. 2) Hyperparameter Tuning manually
Hyperparameter tuning manually is also performed
for the Random Forest Classifier and noted down the results.
The parameters we varied are n_estimators and random_state.
The following hyperparameters and their values are used:
• random_state ->It is used for the shuffling of data to get
good performance of training and testing datasets.
(random_state=42)
• n_estimators ->It is used to represent the number of
decision trees to be used. (n_estimators=700)

IV. ANALYSIS OF RESULTS AND DISCUSSIONS

A confusion matrix is a table structure that can also
be represented visually to summarize the model’s
Fig. 4. Feature Importance of Random Forest Classifier. performance. It can be an efficient tool to evaluate ML models
both correctly and incorrectly. Before Hyperparameter tuning,
Since entropy has the least feature importance and the accuracy score was 99.27%. After Hyperparameter tuning
contributes the least with respect to the model, we drop the using GridSearchCV, the accuracy increased from 99.27% to
feature entropy. With the remaining features, the Random 99.4%. Based on the results obtained from Feature Importance,
Forest Classifier algorithm is run again and the accuracy we observed that entropy contributes the least. Therefore,
jumps from 99.27% to 99.64% upon considering the dropping the entropy feature the accuracy obtained is 99.64%.
remaining features.
Once the Random Forest Classifier is tuned
E. Hyperparameter Tuning manually using the parameters as discussed, the accuracy has
There are different Machine Learning algorithms and increased to 99.99% which is the highest accuracy obtained
each of them differs in many ways. One way these algorithms using Random Forest Classifier. The confusion matrix after
differ is the input parameters which are called manual parameter tuning is depicted in Fig. 5.
hyperparameters. Changing the values of these parameters
affects the output of the model. The steps involved in
obtaining the best values in order to get the best output are
called Hyperparameter Tuning. Hyperparameter tuning on
Random Forest Classifier is done to obtain better results. From
the above comparative study of the six algorithms’ accuracy
scores, we can say that the Random Forest Classifier has the
best accuracy score of 99.27%. We perform Hyperparameter
tuning for Random Forest Classifier to improve the prediction
and model training phase.
Fig. 5. Confusion matrix after tuning the parameters manually.
1) Hyperparameter Tuning using GridSearchCV
GridSearchCV is used to search through a range of Therefore, by doing the parameter tuning manually, we
parameter values to get the best combination of parameter have achieved the Random Forest Classifier's greatest level of
accuracy.

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 09,2025 at 11:04:26 UTC from IEEE Xplore. Restrictions apply.
 we have considered LIME and SHAP as the algorithms for our
dataset. The other approaches include ELI5. XAI is
demonstrated for the model that we obtained by tuning the
parameters manually because this model has got the highest
accuracy. The data set is split in the ratio of 80:20.
A. LIME
LIME(Local Interpretable Model-agnostic
Explanations) is an explainable AI technique that helps us in
explaining and predicting the outcomes of ML algorithms. It
is also used to overcome the drawbacks of ML algorithms like
their “Black Box'' nature. We have used a popular XAI
approach called LIME which provides interpretable and
Fig. 6. ROC Curve for Random Forest Classifier explainable insights into the factors that make an impact on
the model’s prediction. Prediction is done using local
Since the Random Forest Classifier has the best approximations. The Python LIME package makes this task
accuracy, an additional metric, ROC Curve, is implemented even easier. The LIME approach provides a visual
for the above-mentioned algorithm and it is depicted in Fig. 6. representation of the factors that contributed to the model’s
prediction. Here, we have considered the data sample with
TABLE II. Different method accuracy Comparison of Random Forest
Classifier
1.5691(variance), 6.3465(skewness), -0.1828(curtosis), -
Methods Accuracy 2.4099(entropy) and class=0 for prediction. The prediction
Without Tuning 99.27 probabilities and the reason behind the predictions are shown
GridSearchCV 99.44 in Fig. 7. The value 0 indicates a real currency and 1 indicates
Using Feature Importance 99.64 a fake. From Fig. 7 there is 1.00 i.e., 100% confidence that the
Manual Tuning 99.99 first sample is a real currency and 0% confidence that it will
be a fake currency due to the impact of features. The features
As discussed, the Random Forest Classifier has the curtosis, entropy, and skewness contribute positively to the
highest accuracy when compared to the other algorithms used. sample being fake and variance contributes negatively. Since
Therefore we chose this algorithm for further analysis where the variance contributes more negatively, the data point is
we performed hyperparameter tuning using GridSearchCV classified to 0 which is shown using LIME, and is exactly
and manually as well. Accuracy comparison of Random what the target variable class holds for the data point (class)=0.
Forest Classifier using different mentioned methods is
depicted in Table II. The manual tuning had the highest
accuracy and hence Random Forest Classifier with manual
parameter tuning gave the best results in terms of accuracy.
The accuracy levels that were obtained utilizing various
techniques in the studies to which we have referenced are
shown in Table III below. It contains the paper number, the
technique employed, and the accuracy discovered.
TABLE III. ACCURACY COMPARISON OF THE EXISTING METHOD
Paper Reference Methodology Accuracy
with author used (in %)
name
Ballado[6] Image Processing and 96.67
Canny Edge
Sahana Image Processing 90
Murthy[8]
Fig. 7. Explainable AI using LIME
Aman Bhatia[13] Machine Learning 99.9
Algorithm-KNN B. SHAP
Our Proposed Machine Learning 99.99 For binary classification problems, SHAP(SHapley
Method Algorithm - Random
Forest Classifier
Additive exPlanations)plots yield a more illuminating result.
It can see the observation's ground truth label, the model's
forecast for the same observation, a comparison statement that
V. EXPLAINABLE AI (XAI) determines if the model properly predicted this observation
Explainable AI(XAI) is a set of procedures and steps and the force plot(s) that illustrate the model's output for that
that allows human beings to understand the outputs given by observation.
ML algorithms. It is used to describe the AI model and its
impact. With the help of AI explainability, an organization can
take a responsible stance toward the development of AI. As
and when AI becomes advanced, it becomes impossible to
interpret how algorithms came to a result and if the result Fig. 8. SHAP Force Plot
obtained can be trusted or not. Developers can use
explainability to make sure the system is operating as intended. The Force Plot in Fig. 8 shows precisely which attribute
It can monitor, track, explain, and visualize the model. Here, had the greatest influence on the model's forecast for one

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 09,2025 at 11:04:26 UTC from IEEE Xplore. Restrictions apply.
observation in the test sample. The data sample considered is Mrs.Snigdha Sen, Assistant Professor, Department of CSE,
having 3.4663(variance), 1.1112(skewness), curtosis(1.7425), Global Academy of Technology.
and 1.3388(entropy) which is classified as 0(Real).In Fig. 8,
bold 0.12 is the score of the model for the data sample. The REFERENCES
model predicts 1 for higher scores(Fake) and 0(Real) for lower
[1] Amrita, I., Sen, S., & Ashwini, K. (2022). Analysis of Trend in
scores. 0.12 is closer to 0, therefore the sample is classified as COVID-19, World Vaccination and Its Side Effects Using Machine
real. Features that pushed the model to score higher are shown Learning. In Advances in Distributed Computing and Machine
in red, while those that pushed it to score lower are shown in Learning (pp. 317-326). Springer, Singapore.
blue. The features with the greatest influence on the score are [2] Sen, S., Agarwal, S., Chakraborty, P., & Singh, K. P. (2022).
Astronomical big data processing using machine learning: A
those that are situated nearer the line splitting the colors. The comprehensive review. Experimental Astronomy, 1-43.
impact size is depicted by the bar size. Therefore, from the [3] V. Sharan, A. Kaur and P. Singh, "Identification of Counterfeit Indian
above figure, variance has the largest bar, blue in color, and is Currency Note using Image Processing and Machine Learning
located near the boundary, this data sample is ultimately Classifiers," 2023 Third International Conference on Artificial
classified as 0, a real currency. The data point is classified Intelligence and Smart Energy (ICAIS), Coimbatore, India, 2023, pp.
1314-1319, doi: 10.1109/ICAIS56108.2023.10073787.
correctly to 0 which is exactly what the target variable class
[4] P. Garkoti, P. Mishra, N. Rakesh, Payal, M. Kaur and P. Nand, "Indian
holds for the data point(class=0). Currency Recognition System Using Image Processing Techniques,"
2022 9th International Conference on Computing for Sustainable
Global Development (INDIACom), New Delhi, India, 2022, pp. 628-
631, doi: 10.23919/INDIACom54597.2022.9763134.
[5] P. Ponishjino, K. Antony, S. Kumar and S. JebaKumar, "Bogus
currency authorization using HSV techniques," 2017 International
conference of Electronics, Communication and Aerospace Technology
(ICECA), Coimbatore, India, 2017, pp. 179-183, doi:
10.1109/ICECA.2017.8203667.
[6] A. H. Ballado et al., "Philippine currency paper bill counterfeit
detection through image processing using Canny Edge Technology,"
Fig. 9. SHAP Summary Plot 2015 International Conference on Humanoid, Nanotechnology,
Information Technology,Communication and Control, Environment
and Management (HNICEM), Cebu, Philippines, 2015, pp. 1-4, doi:
Each point in Fig. 9 depicts one SHAP value(the average 10.1109/HNICEM.2015.7393184.
marginal contribution of a feature value in all potential [7] Z. Ahmed, S. Yasmin, M. Nahidul Islam and R. U. Ahmed, "Image
coalitions) for a prediction and feature. Red and blue colors processing based Feature extraction of Bangladeshi banknotes," The
indicate a feature's value as higher and lower respectively. A 8th International Conference on Software, Knowledge, Information
generalization of the directionality influence of the attributes Management and Applications (SKIMA 2014), Dhaka, Bangladesh,
may be made based on the distribution of the red and blue dots. 2014, pp. 1-8, doi: 10.1109/SKIMA.2014.7083521.
From Fig. 9, we can conclude that for the dataset as a whole, [8] S. Murthy, J. Kurumathur and B. R. Reddy, "Design and
implementation of paper currency recognition with counterfeit
if the variance is higher, there are chances of the currency detection," 2016 Online International Conference on Green
being real. And entropy is the feature that contributes the least Engineering and Technologies (IC-GET), Coimbatore, India, 2016, pp.
to the model’s prediction. The force plot and summary plot are 1-6, doi: 10.1109/GET.2016.7916838.
used to see the feature dependence of features on a single data [9] Alekhya, D., G. Devi Surya Prabha, and G. Venkata Durga Rao. "Fake
point of test data and all the test set’s data points respectively. currency detection using image processing and other standard
methods." International Journal of Research in Computer and
Communication Technology 3.1 (2014): 128-131.
VI. CONCLUSION
[10] Agasti, Tushar, et al. "Fake currency detection using image
A country’s economic status can be negatively impacted processing." IOP Conference Series: Materials Science and
because of the existence of Fake currency, making fake Engineering. Vol. 263. No. 5. IOP Publishing, 2017.
currency detection crucial and critical. In this paper, we have [11] R. M. Colaco, R. Fernandes, Reenaz and S. S, "Efficient Image
demonstrated an effective method to detect fake currency by Processing Technique for Authentication of Indian Paper Currency,"
2021 International Conference on Computer Communication and
selecting the best model (Random Forest Classifier)out of the Informatics (ICCCI), Coimbatore, India, 2021, pp. 1-8, doi:
6 models taken into consideration based on accuracy and 10.1109/ICCCI50826.2021.9402428.
tuning its hyperparameters using GridSearchCV and manually [12] S. R. Darade and G. R. Gidveer, "Automatic recognition of fake Indian
to increase its accuracy. We have also ensured that the model currency note," 2016 International Conference on Electrical Power and
we have built is working as expected, is accurate in prediction, Energy Systems (ICEPES), Bhopal, India, 2016, pp. 290-294, doi:
can be trusted, and that the features are contributing to the 10.1109/ICEPES.2016.7915945.
correct prediction using LIME and SHAP. Detecting fake [13] A. Bhatia, V. Kedia, A. Shroff, M. Kumar, B. K. Shah and Aryan,
"Fake Currency Detection with Machine Learning Algorithm and
currency will be useful for society and the nation. Choosing Image Processing," 2021 5th International Conference on Intelligent
among the available models and testing them to get the Computing and Control Systems (ICICCS), Madurai, India, 2021, pp.
expected results accurately may be tedious and can involve 755-760, doi: 10.1109/ICICCS51141.2021.9432274.
redoing the tasks. Therefore, this study can be used by [14] M. Jadhav, Y. k. Sharma and G. M. Bhandari, "Currency Identification
researchers to build effective models. As a future and Forged Banknote Detection using Deep Learning," 2019
enhancement of this work, the Image dataset can be used to International Conference on Innovative Trends and Advances in
Engineering and Technology (ICITAET), Shegoaon, India, 2019, pp.
build a robust model. 178-183, doi: 10.1109/ICITAET47105.2019.9170225.
ACKNOWLEDGMENT [15] Dua, D. and Graff, C. (2019). UCI Machine Learning Repository
[http://archive.ics.uci.edu/ml]. Irvine, CA: University of California,
We are very thankful for the constant and immense School of Information and Computer Science.
support, encouragement, and guidance from our guide

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 09,2025 at 11:04:26 UTC from IEEE Xplore. Restrictions apply.