This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

ADate of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Ensemble Deep Learning Based

Prediction of Fraudulent Cryptocurrency
Transactions
QASIM UMER1,2 , JIAN-WEI LI3,4,∗ , MUHAMMAD REHAN ASHRAF2 , RAB NAWAZ BASHIR2 , and
HAMID GHOUS5
1
Department of Computer Science, Hanyang University, Seoul 04763, Korea.
2
Department of Computer Sciences, COMSATS University Islamabad, Vehari 61000, Pakistan. (e-mail: [email protected],
[email protected], [email protected])
3
School of Marxism, Zhejiang Gongshang University, Hangzhou 310018, China. (e-mail: [email protected] )
4
Institute of Education, Xiamen University, Xiamen 361005, China.
5
Department of Computer Science, Institute of Southern Punjab, Multan 60000, Pakistan. (e-mail: [email protected])
Corresponding author: Jian-wei Li (e-mail: [email protected]).

ABSTRACT Cryptocurrency has emerged as a decentralized transaction to overcome the

problems of the centralized transaction system. Although it has become a popular trend
in online cryptocurrency transactions and mobile wallets, this method has increased the
number of fraudulent transactions instead of physically transferring money. Because the
shared data and the history of online transactions may lead to fraudulent transactions.
The preprocess identification of fraudulent cryptocurrency transactions is becoming an
urgent research question. With the exponential blossoming of Artificial Intelligence, the
employing of deep learning in predicting social issues has been achieved in many disciplines.
From this perspective, this paper proposes an ensemble learning approach for fraudulent
cryptocurrency transactions by integrating two deep learning methods: Convolutional Neural
Network (CNN) and Long Short-Term Memory (LSTM). The off-the-shelf CNN and LSTM,
ensemble CNN, and ensemble LSTM with the bagged and boosted approach are compared
in terms of accuracy and losses from training and test datasets. Moreover, the 10-fold
cross-validation approach is employed for the evaluation of the proposed approach. The
evaluation results indicate that the bagged LSTM ensembled approach is significant with
96.4% accuracy and outperforms the other approaches.

INDEX TERMS Cryptocurrency, Convolutional Neural Networks, Long Short Term Memory,
Classification, Blockchain

I. INTRODUCTION have a market cap of more than one billion dollars

[5]. The top ten cryptocurrencies by market cap are
Cryptocurrency has emerged as an exciting plat- shown in Table 1.
form with the potential to overcome problems as-
sociated with the existing modes of payments and Blockchains are used for the development of cryp-
transactions [1]. The tremendous increase in the tocurrencies [6] that maintain public ledgers for
use of cryptocurrency in the payment area has not managing cryptocurrency transactions [7]. Cryp-
only unlocked more opportunities and challenges tocurrency transactions are decentralized and
but involved criminal activities [2]–[4]. According recorded in a peer-to-peer network called a
to one estimate, a thousand cryptocurrencies enter blockchain [8], [9], eliminating the need for a central
the market each month with different usability [5]. authority [10]. Bitcoin is a pioneering cryptocur-
Moreover, more than 12000 cryptocurrencies will be rency, but there are also many other coins with
available by 2022, and 70 of these cryptocurrencies significant potential. For example, Ethereum is the

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

TABLE 1: Market Cap of Top Five Cryptocurrencies

[13]
Cryptocurrency Market Cap
Bitcoin (BTC) $415,203,289,281
Ethereum (ETH) $148,730,510,731
Tether (USDT) $65,953,462,664
USD Coin (USDC) $55,557,354,390
Binance Coin (BNB) $39,260,543,387

second-largest cryptocurrency by market capital-

ization [6] that allows smart contracts [11]. An FIGURE 1: An Overview of Ethereum’s Value [5]
overview of Ethereum’s value is presented in Fig.
1. The Bitcoin network has a limited capacity to
handle many transactions quickly; therefore, it is not is an ideal candidate for this purpose. Many efforts
scalable [10]. Ethereum reduces the problem of the were put into effect using machine learning to detect
scalability of Bitcoin [10]. Vitalik Buterin develops anomalous activity from a different perspective [24]–
Ethereum to delegate power to the user [10]. The [28].
main advantage of Ethereum is its low transaction The purpose of the proposed solutions is to detect
fee (gas) [12] required to execute a transaction fraudulent transactions using a machine learning
on Ethereum, regardless of transaction success or model. The study aims to detect fraudulent trans-
failure. Each gwei (Ethereum gas unit) is equal to actions on the Ethereum platform with limited fea-
0.000000001 ETH (10−9 ETH). Ethereum is more tures. Note that we select the Ethereum platform
adaptable to smart contracts and transactions [1]. because it is widely acceptable and more adaptable
Smart contracts are a type of Ethereum account. to smart contracts. Machine learning approaches are
This means they have a balance and can be the widely applied to improve the accuracy of identifying
target of transactions. Therefore, fraudulent trans- fraudulent transactions. Out of these approaches,
actions may occur through smart contracts. Artificial Neural Network (ANN)-based approaches
The decentralized blockchain approach allows op- [29] are more accurate. The study wants to improve
erations without central control and intermediaries the accuracy of detecting fraudulent transactions in
with several benefits related to privacy and security Ethereum using ensemble deep learning approaches.
[14], e.g., transaction anonymity [15]. Such benefits The study also aims to explore the performance of
make fraudulent behavior very common [16], i.e., different ensemble models to detect the occurrence of
the decentralized control of blockchain and transac- fraudulent transactions using ensemble deep learn-
tions’ anonymity leads to frequent fraudulent trans- ing models.
action behavior in cryptocurrency [17], [18]. Accord- The paper highlights are as follows:
ing to CipherTrace, a cryptocurrency forensics com- • Identification of the fraudulent transactions on
pany’s scam led to a loss of 4.5 billion dollars in the Ethereum network with high accuracy.
2019. Moreover, the cryptocurrency monitoring com- • Introduction of an ensemble machine learning
panies declared that Ethereum is the foremost choice approach to improve the accuracy of identifica-
for fraudulent transactions [6]. Although a user’s tion of fraudulent transactions on the Ethereum
anonymity is very suitable for fraudulent transac- network.
tions with any Cryptocurrency network [19], [20], • In-depth comparison of different machine learn-
the lack of control by an authority and anonymity ing models against the proposed ensemble ap-
is very attractive for fraudulent activity [21]. proach of identification of fraudulent transac-
The Ponzi scheme is one of the most vibrant scams tions.
associated with cryptocurrency. The masquerading The rest of the paper is organized as follows. A
schemes are common on the Ethereum network [22]. literature review is given in Section II. The material
Because of immutability and user anonymity, fraud- and method are given in Section III. Section IV is
ulent transactions are difficult to reverse, thus mak- for results and discussion. Finally, the concluding
ing them very attractive for fraudulent transactions remarks and future work is discussed at the end.
[23]. It is also very difficult and time-consuming to
manually sort for fraudulent transactions. Such a II. RELATED WORK
huge set of transactions makes detecting fraudulent H. H. Sun Yin et al. [30] proposed supervised ma-
transactions nearly impossible. The problem is also chine learning-based anomaly and criminal activity
hard regarding time and other resources required detection in a Bitcoin-based ecosystem. The proposed
to detect abnormal activities [23]. Machine learning solution is based on a dataset of 395 million trans-
2

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

actions with 957 unique clusters. The study also coin system. The proposed simulation model helps to
compares the performance of Random Forests (RF), reduce double-spending risks with absolute accuracy.
Decision Trees (DT), K-Nearest Neighbour (KNN), Patrick Monamo [41] recommended Unsupervised
Gradient Boosting, Ada Boosting and Bagging clas- Anomaly Detection, which finds outliers in trends
sifiers for performance detection of anomalies in using trimmed k-means. This portion is compared
Bitcoin-based systems. The Gradient Boosting is the with the dataset of known fraud. Experimental re-
most accurate out of the seven tested models, with sults show that the proposed approach achieves pre-
80.83% accuracy. cision performance on benchmark datasets.
Many researchers [1]–[4], [31] proposed machine S. Sayadi [42] proposed a solution to detect fraud-
learning-based detection of abnormal activity de- ulent transactions in cryptocurrency and proposed
tection in the Ethereum network. The study also a technique to determine anomalies in electronic
compares the performance of Naïve Bayes (NB), transactions of Bitcoin by machine learning, with
Multilayer Perception (MLP), Support Vector Ma- high accuracy with k-mean and SVM. B. Chen et al.
chine (SVM), RF, and KNN. The SVM and RF are [43] proposed machine learning-assisted solutions to
the most accurate out of the tested models, with detect Bitcoin theft transactions. The performance of
99.66% accuracy in predicting the abnormal activity five machine learning models (KNN, SVM, RF, Ad-
detection in Ethereum. aBoost, and MLP) is evaluated to identify the theft
J. Wu et al. [32] and K. Ajay et al. [33] proposed transactions in Bitcoin. The study results reveal that
machine learning-based abnormal activity detection the RF performs best, with an F1-value of 95.9%.
from suspicious users in Ethereum platforms. The Kasera [44] proposed an artificial intelligence-based
study also compares the performance of the differ- approach for identifying fraud in cryptocurrency.
ent machine-learning algorithms. Among them, DT This study focuses on how artificial intelligence pro-
and RF are the most significant with accuracies of vides us with an empirical framework to identify
83.66% and 98.93%, respectively. such frauds to ensure more security in the crypto-
Teng Hu et al. [34] proposed LSTM based sphere.
machine-learning approach for detecting anomalies Digital currency has become more popular in this
in smart contracts in the Ethereum network. The era. According to this research, some critical issues
proposed solution shows high precision in identify- of anomalous behavior are also associated with it
ing anomalies in smart contracts in the Ethereum that cause serious problems for cryptocurrencies.
network. R. Tan et al. [35] proposed a graph Convo- Divija et al. [45] proposed ML techniques to detect
lution Neural Network (CNN) to identify ambiguous the anomalous behavior of the crypto-currency by
transactions with 95% accuracy. using the Bitcoin dataset. The proposed solution
Q. Yuan et al. [36] proposed machine learning- is based on the multivariate Gaussian distribution,
based phishing detection on the Ethereum network. 2-phase clustering, and one-class SVM algorithms
The proposed solution detects fraudulent transac- to detect the outliers in the Bitcoin dataset. It is
tions by transaction information with 84.6% accu- identified from the results that the multivariate
racy. Ibrahim et al. [37] proposed an ensemble ma- Gaussian distribution algorithm has more accuracy
chine learning model with high accuracy using the as compared to the other models.
SVM and RF-based models. W. J. Tsaur et al. [38] Pham and Lee [46] focused on detecting the
proposed account attributes and opcode functional- anomaly, especially in the Bitcoin transaction. The
ity with XGBoost to 95% precision to detect illicit study’s main aim is to identify suspicious transac-
transactions. tions and user behavior. The study used k-means,
Z. Chen et al. [39] proposed a solution to identify unsupervised SVM, and clustering machine learning
the Ponzi scheme on the Ethereum network using techniques. Bruno et al. [47] proposed a command
a supervised machine learning approach. The RF- and control approach based approach on the network
based approach can identify 305 out of 394 Ponzi of Bitcoin. In this case, the group of transactions is
schemes from the test dataset with more than made based on the users. In the next step, features
90% probabilistic confidence. Aditya Singh et al. [4] of every transaction’s group are identified to detect
proposed the temporal debiasing method by using their behavior, such as whether they act system-
Graph Neural Network (GNN) approach for fraud atically or not. An algorithm named "OSVM" was
detection in cryptocurrency. The study compared the proposed for this analysis to obtain samples from
performance of different machine learning models users with legal behavior only. In this case, the Zom-
and benchmarked the performance of the proposed bierCoin botnet and Bitcoin blockchain are used to
solution with existing models. conduct the test in a closed environment. According
Wei and Vincent Lee [40] proposed an exploratory to the depicted results, the proposed technique is
simulation model to detect the anomaly in the Bit- more useful in detecting the bots having a low rate
3

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

(a) Off-the-shelf LSTM Architecture (b) Off-the-shelf CNN Architecture

FIGURE 2: Off-the-shelf CNN and LSTM Architectures

of FP (false positive) in the different cases. where, X is the set of features with inputs given in
Steven et al. [48] mainly focused on identifying Eq. 2.
illicit IDs on the blockchain of the Ethereum coin. X = x1 , x2 , . . . , xn (2)
The proposed three machine learning models named
DT, RF, and KNN are used to detect fraud on the The set of features X is used to predict the occur-
Ethereum blockchain. The study shows significant rence of fraudulent transaction y. The study uses
improvement in time measurement with the help of LSTM and CNN with an ensemble approach for
machine learning techniques. predicting the occurrence of fraudulent transactions.
Although the above-discussed studies propose ma-
chine/deep learning-based approaches (i.e., DT [49], A. OFF-THE-SHELF LSTM
RF [50], SVM [51], and ANN [29]) to identify the dif- The Long Short Term Memory (LSTM) architecture
ferent types of abnormalities and anomalies in differ- is shown in Figure 2a. Recurrent Neural Network
ent types of cryptocurrency platforms, performance (RNN) is a deep learning technique that can retain
improvement is required to avoid anomalies in cryp- information from previous input that is impossible
tocurrency transactions. In this perspective, this pa- with shallow ANN. In RNN, the states of the input
per proposes an ensemble machine learning model layer change with each input; therefore, the RNN
with CNN and LSTM to improve accuracy in identi- suffers from long-term memory. LSTM is an exten-
fying fraudulent transactions, which is different and sion of the Recurrent Neural Network (RNN) that
effective in contrast to off-the-shelf CNN and off-the- overcomes the problem of long-term memory. LSTM
shelf LSTM models. Notably, Off-the-shelf CNN and is an RNN-type model that can manage the long-
off-the-shelf LSTM refer to pre-trained CNN that term memory to manage the long terms trends in
are readily available and can be used for various data. For this purpose, the LSTM uses the memory
tasks without the need for extensive training from cells whose state management operations are man-
scratch. We combined LSTM and CNN for the en- aged by the gates. The ability to retain long-term
semble approach because of their specific strengths context is useful for problems that require previous
in processing sequential and structural data, re- contextual data in predictions. The new input data
spectively. Fraudulent cryptocurrency transactions and previous hidden states are fed into the model.
often exhibit patterns that can be captured through The model generates the vectors of the element with
sequential analysis. However, cryptocurrency trans- values in the range of 0 and 1 with the sigmoid
action data can have a structural aspect, i.e., the function. The model is trained so that the forget
relationships between entities. By combining LSTM gate is close to 1 when the output is relevant and 0
and CNN, LSTM can capture temporal patterns when the output is irrelevant. Moreover, LSTM can
and dependencies within individual transactions or maintain long-term memory with memory cells and
sequences of transactions, while CNN can extract gates instead of a hidden layer. The additional state
structural features from the transaction data. Soti at time instant ti is maintained by the output of
the forget gate is mentioned by ft by Eq. 3, where t
III. METHODOLOGY is the timestamp, xt is the input, ht−1 is the previous
The objective of the study is to find a function that hidden state, Wi is the weighted matrix, and bt is the
can determine the occurrence of fraudulent transac- bias. Forget gate fc decides which information must
tions y from a set X of features by Eq. 1. retain and which to ignore.

F (X) → y (1) ft = σ(Wi .[ht−1 , xt ] + bt ) (3)

4

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

(a) Bagged CNN Architecture (b) Bagged LSTM Architecture

FIGURE 3: Bagged CNN and LSTM Architectures

The information from current input xt and hidden Algorithm 1 Bagged CNN and Bagged LSTM
state ht−1 are passed through the sigmoid function 1: procedure B AGGED CNN AND B AGGED LSTM
that produces values between 0-1. The value of ft is 2: Input: X, y, N g
used for point-to-point multiplication. 3: Initialize: h ← 1
The input gate it is used to update the cell status. 4: while h ≤ N g do
The previous hidden state ht−1 and current state
R
5: (Xh , yh ) ← rep(X, y) // generate subset
xt are passed through another sigmoid function to (Xh , yh ) of (X, y) for Rrandom sampling with a
classify inputs into important (1) and non-important replacement function rep
(0). The it function is expressed by Eq. 4, where t is g
(σ,W g ,bg )
R
6: Xh −− h
−−−−h−−h→ yh // Training of the hth
the time stamp, it is the input gate, Wr is the weight
model in an ensemble using the (Xh , yh )
matrix, and bf is the bias vector. Moreover, C̃t is the
7: h←h+1
value by tanh function expressed by Eq. 5, where Wc
8: end while
is the weighted matrix of tanh , bc is the bias vector Rg Rg
9: Output: 1 (σ, W1g , bg1 ), . . . , N (σ, WN gg , bN gg )
concerning Wc .
10: end procedure
where, X is the set of features extracted from
it = σ(Wr .[ht−1 , xt ] + bf ) (4) correlation analysis mentioned in Section III-E, y
is the instance of the occurrence of a fraudulent
transaction, N g is the total number of the model in
C̃t = tanh (Wc .[ht−1 , xt ] + bc ) (5) bagging, h is the current model, Xh is the subset
of feature set X for h model, yh is the R instance
The forget vector ft is multiplied by the previous cell
of fraudulent transaction for h model, rep is the
state Ct−1 . The output of this operation is the point-
replacement function, σ is the non-linear activa-
to-point addition with the output of input vector it
tion
R g function, W Ris the weights, b is the bias, and
to produce a new cell state Ct expressed by Eq. 6. g g g
1
(σ, W 1 , b1 ), . . . , N (σ, WN gg , bN gg ) are the set of mod-
els (CNN and LSTM) in bagging.
Ct = ft × Ct−1+it × C̃t (6)

The output gate Ot is used to determine the next hid-

den state by the current state and previous hidden B. OFF-THE-SHELF CNN
state ht−1 through the sigmoid function expressed Although CNN is commonly used for computer
by Eq. 7. The hidden state ht is used for making vision-based solutions, it is also achieved promising
predictions expressed by Eq. 8. results in other domains. CNN is based on various
layers where the output of each layer is connected
Ot = σ(Wo .[ht−1 , xt ] + bo ) (7) to regions of input features. This operation involves
convolving the input features to model filters for
pattern recognition in the input data. The convo-
ht = Ot × tanh C̃t (8) lution layer is the first type of layer in the CNN
architecture, as shown in Figure 2b.
5

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

(a) Boosted CNN Architecture (b) Boosted LSTM Architecture

FIGURE 4: Boosted CNN and LSTM Architectures

Algorithm 2 Boosted CNN and Boosted LSTM Algorithm 3 Integrated Ensemble Model
1: procedure B OOSTED CNN AND B OOSTED 1: procedure I NTEGRATED E NSEMBLED M ODEL
LSTM 2: Input: XtT +1 , α gb ,
R gb g g R gb g g
2: Input: X, y, N b , αt (σ, Wh , bh )· · · N g (σ, WN g , bN g )
b b b b
h
3: Initialize: h←2 b
b b
Initialize: ŷt+1 , h ← 2
Rb b b
1
(σ,W1 ,b1 ) 3:
4: X −−−−−−−−→ y // Train Initial model (X,y) R t
(σ,W1t ,bt1 )
5: while h ≤ N b do P 4: X −−1−−−−−−→ y
g
h=1 R b 5: while h ≤ N b do
6: ttemp ← y − αb m=1 m (.) g R g g g
Rb
h
(σ,Whb ,bbh ) 6: ŷtT +1 ← ŷtT +1 + α b hb (σ, Wbb , bhb , XtT +1 )
7: X −−−−−−−−→ ytemp // Train Initial model 7: h=h+1
(X,y) 8: end while
8: h=h+1 9: Output: ŷtT +1
9: end while Rb Rb 10: end procedure
10: Output: 1 (σ, W1b , bb1 ), . . . , N (σ, WN bbg , bN bbg ) where, XtT +1 is the feature set at time in-
11: end procedure stances, α gb are Weights of bagging or boosting,
where, X is the set of features extracted from cor- R gb g g Rg g g
(σ, Whb , bhb )· · · Nb g (σ, WN gb , bN gb ) are set of ensem-
relation analysis mentioned in Section III-E, y is h
b
b b

the instance of the occurrence of a fraudulent trans- bled bagged or boosted models, ŷtT +1 is the output of
action, N b is the total number of the models in the ensembled model, X is the feature set, Y is the
boosting, αt are weights in boosting models, h is the instance
g
of the output, α is the activation function,
current model, α is nonlinear Rb activation, W is the and Wbb are Weights of bagging or boosting models.
weights, b is the bias, 1 (σ, W1b , bb1 ) is the hth model,
Rb Rb
and 1 (σ, W1b , bb1 ), . . . , N (σ, WN bbg , bN bbg ) are the set of
models (CNN and LSTM) in boosting.
Gn = f (Mn ; ω1 , ω2 , . . . , ωk ) (11)
Note that we select CNN over ANN (the most signif-
Each convolution layer is defined by Eq .9, Eq. 10, icant state-of-the-art approach) because CNN tends
and Eq. 11, where ϕ is the filter at layer 1 (Mn ), Φ is to be a more powerful and accurate way of solving
the filter at layer 2 (Ln ), and ω is the filter at layer classification problems and high accuracy in weight
3 (Gn ). sharing [52].

C. ENSEMBLE MODELS
Mn = f (In ; ϕ1 , ϕ2 , . . . , ϕk ) (9)
Ensemble machine learning models [53] are the
techniques to combine multiple models for predicting
optimal results by combining their output. An en-
Ln = f (Mn ; Φ1 , Φ2 , . . . , Φk ) (10) semble machine learning multiple models are used
6

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

rather than a single model to improve accuracy and TABLE 2: Configurations

consistency. Using multiple machine learning rather Parameter CNN LSTM
than a single model can produce optimal models.
No hidden layers 2 2
Bagging and boosting [54] are the most common
techniques of ensemble machine learning models. Activation Function at the in- Relu Relu
put layer
Both techniques are applied using CNN and LSTM
Activation Function at hidden Relu Relu
models. Using the bagging and bootstrap aggregat- layer 1
ing, four architectures (bagged CNN, Bagged LSTM,
Activation Function at hidden Relu Relu
Boosted CNN, and Boosted LSTM) of ensemble ma- layer 2
chine learning models are made to predict fraudu-
Activation Function at the out- Sigmoid Sigmoid
lent transactions on the Ethereum platform. put layer
In the case of the bagging technique, multiple
Optimizer Adam Adam
models are trained in parallel on a subset of the
No of Epochs 300 300
dataset. The output of each model is combined to
produce a single output. The architecture of bagging Training test ratio 80:20 80:20
CNN and bagged LSTM is shown in Figure 3 where No Nodes at input layer 46 46
Figure 3a presents the bagged CNN architecture and No Nodes at hidden layer 1 46 46
Figure 3b presents the bagged LSTM architecture. No Nodes at hidden layer 2 30 30
The dataset is partitioned into a subset for multiple
No Nodes at output layer 1 1
models in both architectures, and the output is com-
bined to produce optimal results. The results of each
model are combined as expressed in Eq. 12.
in Ethereum. Note that we select the Ethereum
Z g Z g Z g platform because it is widely acceptable and more
(σ, W1g , bg1 ), (σ, W2g , bg2 ), . . . , (σ, WN gg , bN gg ) adaptable to smart contracts. Second, datasets about
1 2 N
(12) the fraudulent transaction of other cryptocurrencies
The algorithms for bagged CNN and bagged LSTM are not publicly available. We performed the cor-
are given by algorithm 1. relation analysis for the identification of important
In the case of boosting CNN and boosting LSTM, features by dropping the highly correlated feature
the multiple models are trained in sequence, and the ( > 0.8) to escape the curse of dimensionality. The
base model depends upon the previous model’s out- Pearson correlation method is used for the corre-
put. The architecture of boosted CNN and boosted lation analysis where the minimum number of ob-
LSTM is shown in Figure 4 where Figure 4a servations is 1. The correlations between features
presents the boosted CNN architecture and Figure and the occurrence of fraudulent and non-fraudulent
4b presents the boosted LSTM architecture. transactions are shown in Figure 5a and Figure
The algorithms for boosted CNN and boosted 5b, respectively. The dark and light colors of the
LSTM are given by algorithm 2. chart depict the strong and weak relationship be-
The predictions by bagged CNN, bagged LSTM, tween features and the occurrence of fraudulent and
boosted CNN, and boosted LSTM is made by algo- non-fraudulent transactions. Notably, the exploited
rithm 3. dataset is imbalanced [55]. Therefore, we perform re-
sampling to correct the bias in the original dataset.
D. CONFIGURATION OF MODELS We only consider the under-sampling for our experi-
The configuration details of the model are given in ments by randomly selecting the n samples from the
Table 2. A similar configuration is used for individ- majority class, where n is the number of samples
ual and ensemble approaches for each layer. These from the minority class. We avoid over-sampling be-
parameters are optimum with the best accuracy. The cause it involves replicating minority class samples
maximum accuracy is achieved with 300 epochs, and to increase their representation in the dataset and
accuracy above 300 is unaffected. Note that we have can lead to overfitting [56], where the model becomes
tried the off-the-shelf algorithms with multiple pa- overly specialized to the minority class and performs
rameter tuning. However, none of them contributed poorly on unseen data. Moreover, it may artificially
to performance improvement. Therefore, we selected improve performance metrics [57], i.e., accuracy.
ensemble approaches.
IV. EVALUATIONS
E. DATASET The performance of selected models is compared
We collected the public dataset from Kaggle [55] using the accuracy and loss from the training and
and reused it for detecting fraudulent transactions test dataset. The dataset is partitioned into a 80:20
7

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

(a) Correlation between features and the occurrence of (b) Correlation between features and the occurrence of non-
fraudulent transactions fraudulent transactions
FIGURE 5: Correlation between features and the occurrence of fraudulent and non-fraudulent transactions

ratio for the training and testing against each fold. LSTM over 300 epochs is shown in Figure 6b
The training and test accuracy of each model is for both the training and testing datasets, re-
analyzed and compared. spectively. The accuracy of the boosted LSTM is
Accuracy is the measure of correct predictions out less than the bagged LSTM for both the training
of total predictions l expressed by Eq. 13. The accu- and test datasets.
racy of each model is observed over three hundred • The accuracy of off-the-shelf LSTM for both the
epochs. training and testing dataset is shown in Figure
6c. The individual LSTM is also in optimum
Correctpredictions shape. The training and testing accuracy of in-
Accuracy = (13)
T otalpredictions dividual LSTM reaches a maximum of 94.7%
and 90.5% over 300 epochs, respectively. The
The 10-fold cross-validation results of six approaches
training and test accuracy of individual LSTM is
are presented in Figure 6, and the following obser-
less than the bagged LSTM and boosted LSTM.
vations are made:
• The training and testing accuracy for detect-
• The training and testing accuracy for detect- ing fraudulent transactions by the bagged CNN
ing fraudulent transactions by bagged LSTM model is 95% and 94%, respectively. The accu-
model is 96% and 96.4% with training and racy of bagged CNN over 300 epochs is shown in
test datasets. Figure 6 presents the accuracy Figure 6d. The bagged CNN model is also bal-
of LSTM and CNN with bagged, boosted, and anced to make predictions for detecting fraud-
simple LSTM/CNN. The accuracy of the model ulent transactions over time. The accuracy of
from training and testing datasets over 300 the bagged CNN is less than the bagged LSTM,
epochs is displayed in Figure 6a. Figure 6a boosted LSTM, and off-the-shelf LSTM models
shows that the bagged LSTM model is in a bal- from both training and test datasets.
anced shape. Consequently, the bagged LSTM • The accuracy of boosted CNN for training and
ensemble model is more accurate than other testing datasets over 300 epochs is shown in
tested models. Figure 6e. The boosted CNN model is also op-
• The training and testing accuracy (shown in timum fitted to predict the occurrence of fraud-
Figure 6b) for detecting fraudulent transactions ulent transactions over time. The maximum ac-
by boosted LSTM model is 95% and 94.5%, curacy of boosted CNN from training and testing
respectively. The boosted LSTM model is also in datasets is 89% and 86%, respectively. The ac-
a balanced shape. The accuracy of the boosted
8

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

(a) Accuracy of Bagged LSTM (b) Accuracy of Boosted LSTM (c) Accuracy of LSTM

(d) Accuracy of Bagged CNN (e) Accuracy of Boosted CNN (f) Accuracy of CNN
FIGURE 6: Accuracy with bagged, boosted, and off-the-shelf LSTM/CNN

curacy of boosted CNN is less than the bagged

CNN, bagged LSTM, boosted LSTM, and off-the-
shelf LSTM.
• The accuracy of individual CNN over 300 epochs
for training and testing datasets is shown in
Figure 6f. The maximum accuracy of the CNN
model is 85.7% and 85.6%, respectively. The
accuracy of the off-the-shelf CNN is less than
all selected models for detecting fraudulent
transactions in Ethereum-based cryptocurrency (a) Comparison of training accuracy
transactions. of selected models
Moreover, Figure 7 presents the comparison of
training and testing accuracy of all the selected mod-
els and The maximum accuracy achieved by each
model, respectively. From Figure 7, the following
observations are made:
• The accuracy of the training dataset of each se-
lected model is shown in Figure 7a for compari-
son purposes. From Figure 7a, it is clear that the
bagged LSTM offers high accuracy as compared
(b) Comparison of testing accuracy
to other selected models. The accuracy of CNN of selected models
is less than all of the selected models. It is also
important to observe that the accuracy of LSTM FIGURE 7: Comparison of training/testing accuracy
is more than the bagged CNN and boosted CNN. of selected models
The accuracy of each model from the testing
dataset is shown in Figure 7b for comparison
purposes. The bagged LSTM is more accurate Ethereum-based transactions.
than other models with the testing dataset. The • The maximum accuracy for detecting fraudu-
performance of the LSTM in terms of accuracy lent transactions is achieved through the bagged
is better than the ensemble CNN and individual LSTM model. The accuracy of the bagged LSTM
CNN for detecting fraudulent transactions in model is 96% through the training dataset and
9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

(a) Loss of Bagged LSTM (b) Loss of Boosted LSTM (c) Loss of LSTM

(d) Loss of Bagged CNN (e) Loss of Boosted CNN (f) Loss of CNN
FIGURE 8: Losses with bagged, boosted, and simple LSTM/CNN

96.4% through the testing dataset. The CNN

model is the least accurate compared to other
models, with the training and testing dataset,
with 85.7% and 85.7% accuracy, respectively.
Note that bagging decreases variance, not bias,
and solves over-fitting issues in a model. There-
fore, the proposed model significantly improves
the performance of the classifier.
Finally, the loss measures the difference between
the predicted and actual values expressed by Eq. (a) Comparison of training log loss
14. Loss is the measure of the difference between of selected models
predicted and actual values. Although the loss is the
inverse function of accuracy, it helps the researcher-
s/readers better understand the evaluation results.
Therefore, we compute the losses of the selected
models in this paper.

Loss = abs(P redictedV alue − ActualV alue) (14)

It is a binary classification problem to detect the

occurrence of a fraudulent transaction or not. The
(b) Comparison of testing log loss of
loss for the problem is measured in the form of selected models
binary cross-entropy, also named log loss. The binary
cross-entropy is the negative average of the log of FIGURE 9: Comparison of training/testing log losses
corrected predicted probabilities expressed by Eq. of selected models
15.

N
of the occurrence of non-fraudulent transactions. The
1 X log loss of selected models is analyzed against the
LogLoss = −(yi × log(pi) + (1 − yi ) × log(1 − pi))
N i=0 training and test datasets.
(15) Figure 8 presents the losses of LSTM and CNN
Where, pi is the probability of the occurrence of a with bagged, boosted, and simple LSTM/CNN. The
fraudulent transaction and (1-pi) is the probability log losses of bagged LSTM for training and testing
10

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

datasets over three hundred epochs are shown in The generalization of the proposed approach is
Figure 8a. The minimum log losses with the training a matter of concern in terms of external validity.
and testing dataset are 0.0273 and 0.0274, respec- Our analysis focuses exclusively on fraudulent trans-
tively. The loss with the bagged LSTM ensemble actions involving Ethereum, and the performance
model is less than the other tested models. The of the approach may differ when applied to the
log losses of boosted LSTM for training and testing prediction of fraudulent transactions in other cryp-
datasets over three hundred epoch is shown in Fig- tocurrencies.
ure 8b. The minimum log losses with the training Additionally, the limited number of fraudulent
and testing dataset are 0.017 and 0.017, respec- transactions poses a threat to external validity. Deep
tively. The log losses of individual LSTM for training learning algorithms typically require fine-tuning pa-
and testing datasets over three hundred epochs are rameters and a substantial amount of training data
shown in Figure 8c. The minimum log losses with the to achieve optimal performance. The scarcity of
training and testing dataset are 0.0273 and 0.0274, fraudulent transactions may restrict the applica-
respectively. bility of our results and hinder a comprehensive
The log losses of bagged CNN for training and exploration of the parameter space.
testing datasets over three hundred epoch is shown
in Figure 8d. The minimum log losses with the V. CONCLUSION
training and testing dataset are 0.031 and 0.00325, The study proposed ensemble deep machine learning
respectively. The log losses of boosted CNN for train- models to predict the occurrence of fraudulent trans-
ing and testing datasets over three hundred epochs actions in the Ethereum network to improve the
are shown in Figure 8e. The minimum log losses predicting accuracy of fraudulent transactions. For
with the training and testing dataset are 0.032 and this purpose, off-the-shelf CNN, off-the-shelf LSTM,
0.033, respectively. The log losses of individual CNN and ensemble approaches are tested and compared
for training and testing datasets over three hundred for performance. For ensemble deep learning, the
epoch is shown in Figure 8f. The minimum log losses bagged and boosted approach is used. Out of the
with the training and testing dataset are 0.0421 six models (off-the-shelf CNN, off-the-shelf LSTM,
and 0.043, respectively. The individual CNN is least bagged CNN, bagged LSTM, boosted CNN, and
efficient in reducing the losses with training and boosted LSTM), the bagged LSTM is more accurate
test datasets compared to others selected models for compared to other models evaluated for the com-
comparison purposes. parison approach. Moreover, the bagged LSTM ap-
Figure 9 compares all the selected models’ training proach is 2.4% more accurate than the state-of-the-
and testing losses. The losses of each model with the art approach for detecting fraudulent transactions
training and testing dataset are shown in Figure 9a in Ethereum. In future, we would like to investigate
and Figure 9b, respectively. It can be observed that the practical applications of the proposed approach
the bagged LSTM ensemble model is more efficient by implementing it in blockchain technology using
in reducing losses than the other models. The off-the- smart contracts.
shelf CNN model is less efficient in reducing losses
than the other models. The bagged LSTM ensemble ACKNOWLEDGMENT
model is more efficient in reducing losses than the This research was supported by the National Nat-
other model. ural Science Foundation of China (62177042); Na-
tional Social Science Foundation, 2022 “Research on
The bagged LSTM model is more accurate com-
The Basic Connotation, Evaluation Criteria and En-
pared to the selected models. The state-of-the-
dogenous and Exogenous Cultivation Mechanism of
art approach for detecting fraudulent transactions
‘Great Masters’ of Ideological and Political Courses
achieved an accuracy of 94% [6]. The proposed
in the New Era” (22VSZ099); and Institute of In-
ensemble learning model achieved an accuracy of
formation & Communications Technology Planning
96.4%. The proposed ensemble LSTM model is also
& Evaluation (IITP) grant funded by the Korea
more efficient in reducing error than the state-of-the-
government (MSIT) (No. 2021-0-00590, Decentral-
art approach.
ized High Performance Consensus for Large-Scale
Blockchains).
A. THREATS TO VALIDITY
The implementation of the proposed approach gives REFERENCES
rise to concerns about internal validity. To ensure [1] M. J. Shayegan and H. R. Sabor, “A collective anomaly detection
its accuracy, we perform cross-checks; however, it is method over bitcoin network,” arXiv preprint arXiv:2107.00925,
2021.
possible that some errors may have been uninten- [2] B. Podgorelec, M. Turkanović, and S. Karakatič, “A machine
tionally overlooked. learning-based method for automated blockchain transaction

11

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

signing including personalized anomaly detection,” Sensors, Conference on Blockchain (Blockchain). IEEE, 2019, pp. 266–
vol. 20, no. 1, p. 147, 2019. 273.
[3] P. K. Choudhary, “Fraudulent account recognition using super- [23] R. M. Bratspies, “Cryptocurrency and the myth of the trustless
vised learning in ethereum,” Ph.D. dissertation, Indian Institute transaction,” Mich. Tech. L. Rev., vol. 25, p. 1, 2018.
of Technology Jodhpur, 2021. [24] J. Cárdenas-Rodríguez, M. Restrepo Sierra, and D. Plazas Es-
[4] A. Singh, A. Gupta, H. Wadhwa, S. Asthana, and A. Arora, “Tem- cudero, “Cryptocurrency scams,” 01 2018.
poral debiasing using adversarial loss based gnn architecture [25] E. Jung, M. Le Tilly, A. Gehani, and Y. Ge, “Data mining-
for crypto fraud detection,” in 2021 20th IEEE International based ethereum fraud detection,” in 2019 IEEE International
Conference on Machine Learning and Applications (ICMLA). Conference on Blockchain (Blockchain). IEEE, 2019, pp. 266–
IEEE, 2021, pp. 391–396. 273.
[5] C. Prices, “Charts and market capitalizations| coinmarket- [26] R. Tan, Q. Tan, P. Zhang, and Z. Li, “Graph neural network
cap.(sem data).” for ethereum fraud detection,” Proceedings - 12th IEEE Inter-
[6] K. Lašas, G. Kasputytė, R. Užupytė, and T. Krilavičius, “Fraudu- national Conference on Big Knowledge, ICBK 2021, pp. 78–85,
lent behaviour identification in ethereum blockchain,” in CEUR 2021.
Workshop Proceedings [Electronic Resource]: IVUS 2020, Infor- [27] R. M. Aziz, M. F. Baluch, S. Patel, and A. H. Ganie,
mation Society and University Studies, Kaunas, Lithuania, 23 “Lgbm: a machine learning approach for ethereum fraud
April, 2020: Proceedings. Aachen: CEUR-WS, vol. 2698, 2020. detection,” International Journal of Information Technology
[7] J. Hamrick, F. Rouhi, A. Mukherjee, A. Feder, N. Gandal, 2022 14:7, vol. 14, pp. 3321–3331, 1 2022. [Online]. Available:
T. Moore, and M. Vasek, “An examination of the cryptocurrency https://link.springer.com/article/10.1007/s41870-022-00864-6
pump-and-dump ecosystem,” Information Processing & Manage- [28] M. Bhowmik, T. S. S. Chandana, and B. Rudra, “Comparative
ment, vol. 58, no. 4, p. 102506, 2021. study of machine learning algorithms for fraud detection in
[8] U. W. Chohan, “Are cryptocurrencies truly trustless?” in Crypto- blockchain,” Proceedings - 5th International Conference on Com-
finance and Mechanisms of Exchange. Springer, 2019, pp. 77– puting Methodologies and Communication, ICCMC 2021, pp.
89. 539–541, 4 2021.
[9] H. Nghiem, G. Muric, F. Morstatter, and E. Ferrara, “Detecting [29] E. Grossi and M. Buscema, “Introduction to artificial neural
cryptocurrency pump-and-dump frauds using market and social networks,” European journal of gastroenterology hepatology,
signals,” Expert Systems with Applications, vol. 182, p. 115284, vol. 19, pp. 1046–54, 01 2008.
2021. [30] H. H. Sun Yin, K. Langenheldt, M. Harlev, R. R. Mukka-
[10] A. Chauhan, O. P. Malviya, M. Verma, and T. S. Mor, “Blockchain mala, and R. Vatrapu, “Regulating cryptocurrencies: a super-
and scalability,” in 2018 IEEE International Conference on vised machine learning approach to de-anonymizing the bit-
Software Quality, Reliability and Security Companion (QRS-C), coin blockchain,” Journal of Management Information Systems,
2018, pp. 122–128. vol. 36, no. 1, pp. 37–73, 2019.
[11] A. Trozze, J. Kamps, E. A. Akartuna, F. J. Hetzel, B. Kleinberg, [31] “Coinbase - Buy Sell Bitcoin, Ethereum, and more
T. Davies, and S. D. Johnson, “Cryptocurrencies and future with trust,” accessed on: 22-Jul.-2022. [Online]. Available:
financial crime,” Crime Science, vol. 11, no. 1, pp. 1–35, 2022. https://www.coinbase.com/
[12] F. Leal, A. E. Chis, and H. González-Vélez, “Multi-service model [32] J. Wu, Q. Yuan, D. Lin, W. You, W. Chen, C. Chen, and Z. Zheng,
for blockchain networks,” Information Processing & Manage- “Who are the phishers? phishing scam detection on ethereum
ment, vol. 58, no. 3, p. 102525, 2021. via network embedding,” IEEE Transactions on Systems, Man,
[13] “Top 10 Cryptocurrencies Price Analysis | Cointele- and Cybernetics: Systems, 2020.
graph,” accessed on: 22-Jul.-2022. [Online]. Available: [33] A. Kumar, K. Abhishek, P. Nerurkar, M. R. Ghalib, A. Shankar,
https://cointelegraph.com/category/top-10-cryptocurrencies and X. Cheng, “Secure smart contracts for cloud-based manufac-
[14] J. Barna, “Blockchain and cryptocurrencies,” 2022. turing using ethereum blockchain,” Transactions on Emerging
[15] B. B. Zarpelão, R. S. Miani, and M. Rajarajan, “Detection Telecommunications Technologies, vol. 33, no. 4, p. e4129, 2022.
of bitcoin-based botnets using a one-class classifier,” in IFIP [34] T. Hu, X. Liu, T. Chen, X. Zhang, X. Huang, W. Niu, J. Lu,
International Conference on Information Security Theory and K. Zhou, and Y. Liu, “Transaction-based classification and de-
Practice. Springer, 2018, pp. 174–189. tection approach for ethereum smart contract,” Information Pro-
[16] M. Bhowmik, T. S. S. Chandana, and B. Rudra, “Comparative cessing & Management, vol. 58, no. 2, p. 102462, 2021.
study of machine learning algorithms for fraud detection in [35] R. Tan, Q. Tan, P. Zhang, and Z. Li, “Graph neural network for
blockchain,” in 2021 5th International Conference on Computing ethereum fraud detection,” in 2021 IEEE International Confer-
Methodologies and Communication (ICCMC). IEEE, 2021, pp. ence on Big Knowledge (ICBK). IEEE, 2021, pp. 78–85.
539–541. [36] Q. Yuan, B. Huang, J. Zhang, J. Wu, H. Zhang, and X. Zhang,
[17] D. Sanz-Bas, C. del Rosal, S. L. Náñez Alonso, and M. Á. “Detecting phishing scams on ethereum based on transaction
Echarte Fernández, “Cryptocurrencies and fraudulent transac- records,” in 2020 IEEE International Symposium on Circuits
tions: Risks, practices, and legislation for their prevention in and Systems (ISCAS). IEEE, 2020, pp. 1–5.
europe and spain,” Laws, vol. 10, no. 3, p. 57, 2021. [37] R. F. Ibrahim, A. M. Elian, and M. Ababneh, “Illicit account
[18] J. astonguay and S. Stein Smith, “Digital assets and blockchain: detection in the ethereum blockchain using machine learning,”
Hackable, fraudulent, or just misunderstood?” Accounting Per- in 2021 International Conference on Information Technology
spectives, vol. 19, no. 4, pp. 363–387, 2020. (ICIT). IEEE, 2021, pp. 488–493.
[19] L. Chen, J. Peng, Y. Liu, J. Li, F. Xie, and Z. Zheng, “Phishing [38] W.-J. Tsaur, J.-C. Chang, and C.-L. Chen, “A highly secure iot
scams detection in ethereum transaction network,” ACM Trans- firmware update mechanism using blockchain,” Sensors, vol. 22,
actions on Internet Technology (TOIT), vol. 21, no. 1, pp. 1–16, no. 2, p. 530, 2022.
2020. [39] W. Chen, Z. Zheng, E. C.-H. Ngai, P. Zheng, and Y. Zhou,
[20] H. Arimura, M. Soufi, H. Kamezawa, K. Ninomiya, and “Exploiting blockchain data to detect smart ponzi schemes on
M. Yamada, “Radiomics with artificial intelligence for precision ethereum,” IEEE Access, vol. 7, pp. 37 575–37 586, 2019.
medicine in radiation therapy,” Journal of radiation research, [40] V. Lee and H. Wei, “Exploratory simulation models for fraudu-
vol. 60, no. 1, pp. 150–157, 2019. lent detection in bitcoin system,” in 2016 IEEE 11th Conference
[21] A. Luchkin, O. Lukasheva, N. Novikova, V. Melnikov, A. Zy- on Industrial Electronics and Applications (ICIEA). IEEE,
atkova, and E. Yarotskaya, “Cryptocurrencies in the global 2016, pp. 1972–1977.
financial system: problems and ways to overcome them,” in [41] P. Monamo, V. Marivate, and B. Twala, “Unsupervised learning
Russian Conference on Digital Economy and Knowledge Man- for robust bitcoin fraud detection,” in 2016 Information Security
agement (RuDEcK 2020). Atlantis Press, 2020, pp. 423–430. for South Africa (ISSA). IEEE, 2016, pp. 129–134.
[22] E. Jung, M. Le Tilly, A. Gehani, and Y. Ge, “Data mining- [42] S. Sayadi, S. B. Rejeb, and Z. Choukair, “Anomaly detection
based ethereum fraud detection,” in 2019 IEEE International model over blockchain electronic transactions,” in 2019 15th

12

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3310576

Umer et al.: Ensemble Deep Learning Based Prediction of Fraudulent Cryptocurrency Transactions

International Wireless Communications & Mobile Computing LI JIANWEI , associate professor and mas-
Conference (IWCMC). IEEE, 2019, pp. 895–900. ter’s supervisor of the School of Marxism
[43] B. Chen, F. Wei, and C. Gu, “Bitcoin theft detection based on of Zhejiang Gongshang University, and doc-
supervised machine learning algorithms,” Security and Commu- toral candidate of the Education Research
nication Networks, vol. 2021, 2021. Institute of Xiamen University, studied
[44] A. Kasera, “Cryptocurrency frauds,” International Journal psychology during his master’s degree and
of Engineering and Advanced Technology (IJEAT), vol. 9, pedagogy during his doctoral degree. His
no. 6, pp. 261–268, Aug. 2020. [Online]. Available:
research interests include but are not lim-
https://doi.org/10.35940/ijeat.F1391.089620
ited to deep learning for students, teacher
[45] G. D. Arya, K. V. S. Harika, D. V. Rahul, S. Narasimhan,
and A. Ashok, “Analysis of unsupervised learning algorithms development literacy in the digital era, and
for anomaly mining with bitcoin,” in Machine Intelligence and the construction of teacher-student relationships.
Smart Systems. Springer, 2021, pp. 365–373.
[46] T. Pham and S. Lee, “Anomaly detection in the bitcoin
system - a network perspective,” 2016. [Online]. Available:
https://arxiv.org/abs/1611.03942
[47] B. B. Zarpelão, R. S. Miani, and M. Rajarajan, “Detection
of bitcoin-based botnets using a one-class classifier,” in IFIP
International Conference on Information Security Theory and
Practice. Springer, 2018, pp. 174–189.
[48] S. Farrugia, J. Ellul, and G. Azzopardi, “Detection of illicit MUHAMMAD REHAN ASHRAF is an Assis-
accounts over the ethereum blockchain,” Expert Systems with tant professor at the Department of Com-
Applications, vol. 150, p. 113318, 2020. [Online]. Available: puter Sciences, COMSATS University Is-
https://www.sciencedirect.com/science/article/pii/S0957417420301433 lamabad, Vehari Campus, Pakistan. He re-
[49] L. Rokach and O. Maimon, Decision Trees, 01 2005, vol. 6, pp. ceived his M.Sc and M.Phil Degree from
165–192. Quaid-I-Azam University Islamabad, Pak-
[50] L. Breiman, “Random forests,” Machine Learning, vol. 45, pp. istan. He is particularly interested in ma-
5–32, 10 2001.
chine learning, data mining and digital im-
[51] T. Evgeniou and M. Pontil, “Support vector machines: Theory
age processing.
and applications,” vol. 2049, 09 2001, pp. 249–257.
[52] K. O’Shea and R. Nash, “An introduction to convolutional neural
networks,” ArXiv e-prints, 11 2015.
[53] C. Zhang and Y. Ma, Ensemble Machine Learning: Methods
and Applications. Springer Publishing Company, Incorporated,
2012.
[54] T. Khoshgoftaar, J. Van Hulse, and A. Napolitano, “Comparing
boosting and bagging techniques with noisy and imbalanced
data,” IEEE Transactions on Systems, Man, and Cybernetics,
Part A, vol. 41, pp. 552–568, 05 2011.
[55] “Ethereum fraud detection dataset,” ac- RAB NAWAZ BASHIR received his Ph.D.
cessed on: 18 July 2023. [Online]. in Computer Science and MS in Com-
Available: https://www.kaggle.com/datasets/vagifa/ethereum- puter Science from Islamia University Ba-
frauddetection-dataset?resource=download hawalpur, Bahawalpur, Pakistan, in 2015
[56] L. Perez and J. Wang, “The effectiveness of data augmentation and 2021. Currently working as a lecturer
in image classification using deep learning,” 2017. at Comsats University Islamabad, Vehari,
[57] C. C. Aggarwal, A. Hinneburg, and D. A. Keim, “On the sur- Pakistan. His research interests are Inter-
prising behavior of distance metrics in high dimensional space,”
net of Things (IoT) applications in agricul-
in Database Theory — ICDT 2001, J. Van den Bussche and
ture.
V. Vianu, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg,
2001, pp. 420–434.

QASIM UMER is an Assistant Professor

at the Department of Computer Sciences, HAMID GHOUS is the head of research for
COMSATS University Islamabad, Vehari the computer science department at the
Campus, Pakistan. He received BS degree Institute of Southern Punjab (ISP). He is
in Computer Science from Punjab Univer- also leading the vision, linguistic, and ma-
sity, Pakistan in 2006, MS degree in .Net chine learning Lab at ISP. He has authored
Distributed System Development from Uni- and co-authored more than 30 publications
versity of Hull, UK in 2009, MS degree in in the past. His main area of research is
Computer Science from University of Hull, Machine and Deep learning methods.
UK in 2013, and Ph.D. degree from Beijing
Institute of Technology, China. He is particularly interested in
machine learning, data mining, and software maintenance. He
is also interested in developing practical tools to assist software
engineers.

13

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4