You are on page 1of 6

IEEE INFOCOM 2022 - IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS) | 978-1-6654-0926-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/INFOCOMWKSHPS54753.2022.

9797992 IEEE INFOCOM WKSHPS: ICCN 2022: IEEE International Workshop on Intelligent Cloud Computing and Networking

Malicious Transaction Identification in Digital


Currency via Federated Graph Deep Learning
Hanbiao Du∗ , Meng Shen† , Rungeng Sun∗ , Jizhe Jia† , Liehuang Zhu† and Yanlong Zhai†
∗ School
of Computer Science, Beijing Institute of Technology, Beijing, P. R. China
† School of Cyberspace Science and Technology, Beijing Institute of Technology, Beijing, P. R. China
{duhanbiao, shenmeng, sunrungeng, jiajizhe, liehuangz, ylzhai}@bit.edu.cn

Abstract—With the rapid development of digital currencies viruses, money laundering, etc. These malicious activities have
in recent years, their anonymity provides a natural shelter for caused great losses to people’s property and seriously damaged
criminals. This problem resulting in various types of malicious the financial order in cyberspace. Ponzi scheme is a fraudulent
transactions emerge in an endless stream, which seriously endan-
gers the financial order of digital currencies. Many researchers investment operation that creates profits for old investors
have started to focus on this area and have proposed heuristics by acquiring funds from new investors, and eventually the
and feature-based centralized machine learning algorithms to subsequent investors may lose most of their own funds [4].
discover and identify malicious transactions. However, these Blockchain-based forms of Ponzi schemes defrauded over $7
approaches ignore the existence of financial flows between digital million in Bitcoin from September 2, 2013 to September 9,
currency transactions and do not use the important neighborhood
relationships and rich transaction characteristics. In addition, 2014 [5]. Since March 2019, TokenBank trading platform
centralized learning exposes a large amount of transaction feature promised high investment returns to investors and suddenly
data to the risk of leakage, where criminals may trace the actual ran away after three months of operation, victimizing over 1
users using traceability techniques. To address these issues, we million people and involving over $3 billion.
proposes a graph neural network model based on federated Recent studies have proposed a number of methods for
learning named GraphSniffer to identify malicious transactions
in the digital currency market. GraphSniffer leverages federated identifying malicious transactions, with the aim of discovering
learning and graph neural networks to model graph-structured illegal behaviors hidden in digital currencies and regulating
Bitcoin transaction data distributed at different worker nodes, the cryptocurrency market and its internal ecosystem. Some
and transmits the gradients of the local model to the server node researchers have put forward to use data mining technology
for aggregation to update the parameters of the global model. and machine learning algorithm to classify and identify Bit-
GraphSniffer can realize the joint identification and analysis of
malicious transactions while protecting the security of transaction coin transaction data to find malicious behaviors related to
feature data and the privacy of the model. Extensive experiments Ponzi schemes and money laundering [6–8]. In the Ethereum
validate the superiority of the proposed method over the state- network, researchers have used supervised learning algorithms
of-the-art. to detect Ponzi schemes deployed as smart contracts [9].
Index Terms—Malicious detection; Digital currency; Graph All these schemes, however, cannot perform well for the
neural network; Federated learning; Blockchain
reasons given below. First, for example, Bitcoin is a dig-
I. I NTRODUCTION ital currency based on the UTXO model. A new UTXO
(payee) generated by one transaction will be used by an-
Bitcoin, famous for its decentralization and anonymity of other transaction (payer), meaning that fund flow from one
transactions, has attracted a lot of attention from researchers transaction to another. This relationship can further aggregate
and investors worldwide since it was first proposed in 2008 [1]. features between multiple neighbors of each transaction for
According to the data from the well-known blockchain digital malicious transaction identification. The traditional machine
currency statistics website CoinMarketCap [2], as of Decem- learning algorithms deployed in current researches empower
ber 2021, the total market capitalization of the top-ranked decision-making capabilities for identifying malicious trans-
Bitcoin is more than $910 billion. However, while anonymity actions, but most of them are difficult to extend beyond the
protects the privacy of users’ identities and transactions, it also immediate neighborhood. Second, traditional machine learning
provides a natural barrier for criminal behaviors, making the algorithms rely on a centralized data set to train the model,
digital currency market a hotbed of malicious transactions [3]. and the native machine needs to collect and store all the data.
At present, there are various malicious transactions in the This adds to the computational load while greatly increasing
digital currency ecosystem, such as Ponzi schemes, ransom the risk of privacy leakage of transaction feature data [10].
Meng Shen is the corresponding author. The criminals may use traceability techniques to obtain the
This work is partially supported by National Key R&D Program of real identity of the user through the stolen data [11–14].
China under Grant 2020YFB1006101, Beijing Nova Program under Grant In this paper, we propose GraphSniffer, a malicious transac-
Z201100006820006, NSFC Projects under Grants 61972039 and 61872041,
and CCF-Huawei Populus Euphratica Blockchain Foundation under Grant tion identification model based on federated learning and graph
CCF-HuaweiBC2021007. neural network, in order to better exploit the neighborhood

Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on March 22,2024 at 04:39:36 UTC from IEEE Xplore. Restrictions apply.
978-1-6654-0926-1/22/$31.00 ©2022 IEEE
IEEE INFOCOM WKSHPS: ICCN 2022: IEEE International Workshop on Intelligent Cloud Computing and Networking

features among Bitcoin transaction data while avoiding the B. Related Work
risk of data privacy leakage. Unlike traditional centralized In order to protect the security of users’ assets and reduce
machine learning, GraphSniffer models Bitcoin transaction the malicious activities of criminals, identifying malicious
data on the worker side using graph neural network to train the digital currency transactions has become a problem worth
respective local models. Afterwards, the server aggregates the exploring. Currently, researchers have conducted a series of
gradients of the local model in a privacy-preserving manner, studies on specific types of malicious transactions.
jointly updates the parameters of the global model, and sends By investigating 192 Bitcoin scam cases, Vasek et al. [5]
the results back to the worker side. The GraphSniffer not classified these schemes into four types: Ponzi scheme, mining
only effectively utilizes the local and aggregated features scheme, fraudulent wallet and bogus exchange. Bartoletti et
between Bitcoin transactions, but also has the advantage of al. [6] showed the research on Bitcoin Ponzi scheme through
possessing data privacy and security protection, enabling the data mining techniques. Henderson et al. [8] used K-means
joint identification and analysis of malicious digital currency and Role Extraction (RolX) methods to identify users who
transactions in data protection scenario. conducted transactions in an abnormal manner over the Bitcoin
The main contributions of this paper are summarized as
network and found that they were usually related to money
follows:
laundering. To understand the behavior of Bitcoin usage in
• We propose GraphSniffer, a malicious transaction identi-
the Dark Web, Foley et al. [16] proposed several features to
fication model based on graph neural network under the identify illegal Bitcoin addresses, which were used to estimate
federated learning framework. By replacing the central- the amount of illegal activities. Mark et al. [17] developed a
ized machine learning by the secure FL platform, we can dataset for identifying money laundering by using Bitcoin’s
protect the privacy of transaction feature data. publicly available transaction information and deployed graph
• We consider using the flow of funds between digital
neural network to provide anti-money laundering solutions.
currency transactions to construct features, in order to Seeking to reduce illegal activities on the Ethereum network,
discover and leverage the neighbourhood relationship. Chen et al. [9] proposed to use supervised learning algo-
• We demonstrate that the proposed method outperforms
rithms to extract features from user accounts and operation
the state-of-the-art on the Bitcoin transaction graph codes of the smart contracts and then build a classification
dataset with extensive experiments. model to detect Ponzi schemes deployed as smart contracts.
To the best of our knowledge, this is the first attempt to ap- It is estimated that there are more than 400 smart-Ponzi
ply a combination of federated learning techniques and graph schemes on the Ethereum network. In addition, McGinn et
neural network to the digital currency malicious transaction al. [18] demonstrated a forced-directed graphical visualization
identification scenario. The rest of this paper is organized as that can accelerate data exploration and discover unexpected
follows. Section II summarizes the current state of research high-frequency trading patterns (including automated money
for malicious transactions in digital currency. The principle laundering operations) in the Bitcoin network.
and system architecture of our method is presented in Section The limitations of existing methods are that using traditional
III. After evaluating the performance of our method in Section machine learning relies on carefully-selected features, cannot
IV, the conclusion of this study is presented in Section V. effectively utilize the neighbor features of transactions, and
II. BACKGROUND AND R ELATED W ORK centralized learning makes feature data facing the risk of leak-
A. Bitcoin age. The GraphSniffer proposed in this paper is based on the
data privacy protection scenario, which utilizes graph neural
Bitcoin is a decentralized digital currency, which does not
networks to learn the neighbor features between transactions
rely on a trusted entity and is a payment system based on
to jointly analyze and detect digital currency transaction data
cryptography and a peer-to-peer network. Transactions and
stored in different sites.
blocks are the main components of Bitcoin. Bitcoin miners
generate blocks by solving cryptographic puzzles (Proof of III. T HE P ROPOSED M ETHOD
Work). With each block generated, the miner can create and
receive a new Bitcoin reward. The entire network automati- In this section, we introduce the details of the proposed
cally adjusts the difficulty of the cryptographic puzzle so that malicious transaction identification model based on federated
the network generates one new block every 10 minutes on learning and graph neural network, GraphSniffer.
average, and the newly generated block contains the hash of
the previous block, thus linking to form the block chain. A. Overview
The basic structure and value unit of a Bitcoin transaction As shown in Figure 1, there are four main processes in each
is the Unspent Transaction Output (UTXO), which cannot be round of training. First, the server delivers the initial model
divided again [15]. The set of UTXOs of Bitcoin contains parameters to each worker. Then, each worker trains the local
all UTXOs, and the balance of the user’s wallet is the total machine learning model using GNN network based on their
amount of UTXOs available. Each UTXO participates in a local data. After that, the worker sends the gradients trained
Bitcoin transaction as a whole, which prevents the problem of by their respective models to the server. Finally, the server
double spending. aggregates these parameters to update the global model.

Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on March 22,2024 at 04:39:36 UTC from IEEE Xplore. Restrictions apply.
IEEE INFOCOM WKSHPS: ICCN 2022: IEEE International Workshop on Intelligent Cloud Computing and Networking

1.Downloading initial model

Local Data GNN Local Model Worker 1


3.Uploading
local gradients

Global Model
Local Data GNN Local Model Worker 2 Server
4.Aggregating parameters to

...
obtain the global model

5.Sending updated parameters


Local Data GNN Local Model Worker n
2.Training local model by local data

Fig. 1: The Workflow of GraphSniffer

With an intuitive description above, we formally present the Algorithm 1 The learning process of GraphSniffer
construction process of GraphSniffer, as depicted in Algorithm Input: Graph Gi = (Vi , Ei ) and node features {xvi , ∀i∈K} on
1. To begin with, GraphSniffer randomly initializes global worker i; activation function δ; learning rate β
model parameters and sends them to worker nodes (line Output: the GraphSniffer model M
1: Initialize and unified the global model parameters W 0
1). At the local model training part, GraphSniffer generates
2: for each epoch t = 0 to T do
local node embeddings by aggregating multi-hop neighbors’ 3: Stage1:Local training in worker
information using GraphSAGE [19] method (line 5-11) and 4: for i ∈ K in parallel do
then computes the gradient of local model (line 12-13) at each 5: # Calculates local node embeddings
worker. After uploading the gradients to server, GraphSniffer 6: for vi ∈ Vi do
(k)
7: Worker i: calculates HN (v) ←
calculates the global model parameters (line 17) and returns (k−1)
them back to update local parameters (line 19-22). Finally, we 8: Aggregatek (Hu , ∀u∈N (v))
(k)
can obtain the convergent GraphSniffer model. 9: calculates Hv ←
(k−1) (k)
10: δ(W (k) ·Concat(Hv , HN (v) ))
Section III-B will focus on the process of training the local 11: end for
model using GNN network on the worker side. Section III-C 12:
(t)
ŷit ← δ(Hv,i , Wit )
will explain how the server side aggregates all the locally 13: Calculates L(ŷit , yit ) and gets gKi
updated gradients for the purpose of data privacy protection. 14: # Uploads gradient gKi to server
15: end for
16: Stage2:Model aggregation in server
B. Local Training With GNN 17: W t+1 ← W t − β· I1 Ii=1 gKi
P

Suppose that we are given a set of I workers K = 18: Stage3:Update local model
19: Send updated parameters to worker i, ∀i∈K
{K1 , K2 , ..., KI }, and each worker node Ki owns a local
20: for i ∈ K in parallel do
Bitcoin transaction graph dataset Gi = (Vi , Ei ). Vi is the node 21: Worker i: updates local parameters
set, where each node in the graph represents a transaction. 22: end for
Ei is the edge set, where the edges represent the flow of 23: end for
transactions that exist between the two parties. Each node 24: return M
vi ∈Vi is associated with a transaction feature vector xvi and a
label yvi . Table I summarizes the major symbols and notations
in this paper. output, we use a weight matrix W (l) . The worker Ki runs a
After receiving the parameters transmitted by the server, forward computation of a graph convolutional layer as follows:
each worker locally trains a graph neural network model with
(l+1) (l) (l)
its own data Gi = (Vi , Ei ) respectively. Generally, during the Hi = δ(Ai Hi Wi ), (1)
training process, the l-th layer of the GCN takes the node
embedding matrix H (l) and the adjacency matrix A as input. where δ is an activation function (such as ReLU) for all but
In order to update the node embedding matrix to H (l+1) as the output layer. Assume that there are L layers of graph

Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on March 22,2024 at 04:39:36 UTC from IEEE Xplore. Restrictions apply.
IEEE INFOCOM WKSHPS: ICCN 2022: IEEE International Workshop on Intelligent Cloud Computing and Networking

convolutions in total, we write the output of an GCN as Based on the above output and loss function, we can obtain
follows: the gradient from each worker node.
(L) (L) (L) (L)
ŷi = f (Ai , Wi ; Hi ) = sof tmax(Ai Hi Wi ) (2) C. Model Aggregation
In the case of node classification, we use the softmax function After receiving all the gradients from the participating
at the output layer to get the normalized probability. workers, the server leverages these gradients by aggregating
new global parameters. The gradient of each worker Ki is
TABLE I: Symbols and Notations. represented by gKi , and the local model of Ki is represented
Notations Descriptions
by θKi . Then, we define:

K set of workers gKi = OL(θKi ). (6)


Gi graph dataset at worker i
Vi node set of graph Gi
They transmit local model gradients to the server for an
Ei edge set of graph Gi average mean gradient to update the global model as:
vi nodes of node set Vi I
xvi feature vector of node vi
1X
W t+1 = W t − β· gK , (7)
yvi label of node vi I i=1 i
Ai adjacency matrix of graph Gi
Wi model weight of worker i where t specifies the current number of epoch, β is the learning
J the total number of classes rate and W t is the global model parameters.
yit,j label probability of class j at the t-th epoch Each participating worker will download the updated global
ŷit,j predicted label probability of class j at the t-th epoch model parameters to train its local model. Repeating the above
H node embedding vector as the GNN input steps, the worker nodes and server node can collaboratively
N (v) neighborhood sampling function train a global graph neural network model while protecting
gKi gradient of worker i data privacy.
θKi local model of worker i
Wt global model parameters at the t-th epoch IV. P ERFORMANCE E VALUATION
However, GCN requires the input of the entire adjacency In this section, we introduce our dataset, experimental set-
matrix and feature matrix for training, and each node needs tings, methodology and the results to validate the effectiveness
the entire topological graph’s Laplacian matrix to participate of the proposed method.
in the computation when doing convolution. When a new node A. Experimental Settings
comes in, the entire graph nodes have to be involved in the
1) Environments: In our scheme, the training operations are
computation to get the embedding of this node. As the graph
implemented by Python 3.7 and Pytorch 1.10. The program
grows in size, it consumes a lot of device memory and CPU
runs on a PC (AMD Ryzen 7 2700X Eight-Core Processor
memory. Therefore, we use GraphSAGE to solve this problem.
at 3.70GHz and 32 GB RAM) to simulate the whole training
Different from the traditional full graph convolution, Graph-
process.
SAGE controls the number of neighbor nodes participating
2) Dataset: The experimental data chosen for this paper is
in the calculation through random sampling. The sampled
derived from the classical Elliptic Data Set [17], which was
neighbors’ embedding matrices are transmitted to the node,
published by Elliptic Company. To our knowledge, it is the
and an aggregation function is used to aggregate the neighbors’
largest publicly available dataset of labeled transactions of any
information to update the embedding matrix of the node.
digital currency in the world, presented as a graph network of
We use k denotes the current step in the outer loop. First,
Bitcoin transactions.
each node vi ∈Vi aggregates the embedding matrices of the
The Elliptic Data Set is constructed using public informa-
nodes in its immediate neighborhood into a single vector.
tion. The nodes in the graph represent transactions and the
Mathematically, we write:
edges represent the flow of Bitcoin currency (BTC) from
(k)
HN (v) = Aggregatek (Hu(k−1) , ∀u∈N (v)), (3) one transaction to the next (based on the UTXO model of
Bitcoin). There are 203,769 nodes and 234,355 transaction
where N (v) is the neighborhood sampling function. After this,
flows in total. By mapping Bitcoin transactions to real entities
GraphSAGE updates the node’s representation by concatenat-
in licit categories (miners, exchanges, wallet providers, licit
ing the aggregated neighborhood embedding information with
services, etc.) and illicit categories (Ponzi schemes, terrorist
the node’s current embedding information through a nonlinear
organizations, scams, malware, etc.), a transaction is consid-
activation function as:
ered to be licit if a licit entity initiates it. In the Elliptic
(k)
Hv(k) = δ(W (k) ·Concat(Hv(k−1) , HN (v) )), (4) Data Set, 2% (4,545) of transactions are labeled illicit, 21%
Then, the loss function can be defined as follows: (42019) of transactions are labeled licit, and the rest of the
J transactions fall into the category of unknown legitimacy. Each
1 X X t,j node is associated with 166 transaction features, including
L(ŷit , yit ) =− i yi log ŷit,j , (5)
n j=1
94 original features (time step, number of inputs/outputs,

Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on March 22,2024 at 04:39:36 UTC from IEEE Xplore. Restrictions apply.
IEEE INFOCOM WKSHPS: ICCN 2022: IEEE International Workshop on Intelligent Cloud Computing and Networking

transaction fee, etc.) and 72 aggregated features (maximum, TABLE II: Illicit transaction classification results.
minimum, standard deviation, correlation coefficient, etc.).
Illicit Transaction
In order to explore the joint computational analysis in data
Method Accuracy Precision Recall F1-Score
protection scenarios, the dataset used in this paper desensitizes
and splits the Elliptic Data Set to 40 sub-datasets. Each sub- Logistic Regression 0.892 0.677 0.564 0.615
dataset represents a distributed node, which data constitutes Multilayer Perceptron 0.972 0.912 0.675 0.776
a separate graph, and 40 nodes work together to complete GraphSniffer 0.994 0.894 0.751 0.816
federated learning. The pre-processed dataset does not change
the structure of the transaction graph, and the transactions are
labeled into three types (0 is illicit, 1 is licit, 2 is unknown). Training accuracy result in different models
1.0
Each node is associated with 165 transaction features (93
original features and 72 aggregated features), and finally 9

Accuracy of target model


0.9
groups of graph data are used as test data.
3) Metrics: Four widely used evaluation metrics are 0.8
adopted to evaluate our approach and benchmark methods,
including Accuracy, Precision, Recall and F1-Score. The four 0.7
metrics are both the greater the better. We can compute
model precision and recall for illicit transactions by equations 0.6
GraphSniffer
as: Accuracy = (T N + T P )/(T N + T P + F N + F P ), MLP
LR
P recision = (T N )/(T N + F N ), Recall = (T N )/(T N + 0.5
F P ), where T N is the amount of negative instances (illicit 0 20 40 60 80 100 120 140 160 180 200
Epochs
transactions) that are classified correctly, T P is the amount
of positive instances (licit transactions) that are classified Fig. 2: Accuracy result using different models.
correctly, F N is the amount of positive instances (licit trans-
actions) that are classified incorrectly and F P is the amount the ReLU function for the activation process. To prevent
of negative instances that are classified incorrectly. Then we the model from overfitting and to improve the generalization
can obtain model illicit F1-Score by equation: ability, we add a dropout layer to reduce the coupling between
2·P recision·Recall neurons (dropout rate=0.5). Finally, the probability vector of
F1 = (8) the corresponding category is obtained through Log Softmax
P recision + Recall
function. In Graph Neural Network, we adopted GraphSAGE
B. Benchmark Methods as the graph convolutional layer. There are three hidden layers
The benchmark method used in this paper is to biclassify employing ReLU as activation function. The MLP model and
each transaction using supervised machine learning within a GraphSniffer model are trained for 200 epochs by using the
federated learning framework, including Logistic Regression Adam optimizer and a learning rate of 0.001.
and Multilayer Perceptron (MLP). Logistic Regression method We perform parameters tuning-up for all mentioned models
has the advantage of being highly interpretable and is preferred and report the best results based on the validated dataset and
in anti-money laundering scenarios. Each input neuron in MLP evaluate model performance on the test dataset.
receives one data feature, and the hidden layer is processed
using an activation function during forward propagation. The D. Experimental Results Analysis
output is a probability vector corresponding to each category. We evaluate these three models by using all the 165 features
Logistic Regression and MLP are feature-based machine learn- and the results are summarized in Table II. The bottom part
ing methods, which can implement classification problems but
are challenging to extend beyond the immediate neighborhood. F1-Score in different models
0.8
C. Implementation Details 0.7
F1-Score of target model

In our experiments, all models are implemented with py- 0.6


torch framework. The Logistic Regression, Multilayer Percep- 0.5
tron, and Graph Neural Network methods used in this paper
0.4
are all based on the Federated Learning architecture. In the FL
0.3
part, we adopt 40 worker nodes to train local models using
their respective dataset and aggregate the gradients to server 0.2 GraphSniffer
nodes to update the global model parameters. 0.1 MLP
In Logistic Regression, we use the default parameters from 0.0
LR
the scikit-learn Python package. In MLP, the number of input 0 20 40 60 80 100 120 140 160 180 200
Epochs
neurons is 165 (corresponding to 165 transaction features).
Besides, we set a hidden layer with 96 neurons and choose Fig. 3: F1-Score result using different models.

Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on March 22,2024 at 04:39:36 UTC from IEEE Xplore. Restrictions apply.
IEEE INFOCOM WKSHPS: ICCN 2022: IEEE International Workshop on Intelligent Cloud Computing and Networking

of Table II shows the results achieved when we leverage the like to extend our approach to handle more kinds of digital cur-
graph structure of the Bitcoin transaction dataset with the rencies and make improvements in the operational efficiency of
GraphSniffer (3 GraphSAGE layers). the algorithm to contribute to the financial order in cyberspace.
In the scenario of malicious transaction identification, a
R EFERENCES
small number of illicit ones is more important to be solved.
[1] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” Decen-
Due to the task is a binary classification and the two classes are tralized Business Review, p. 21260, 2008.
imbalanced, we use a weighted cross entropy loss to train the [2] “Coinmarketcap,” Accessed: Dec. 19, 2021. [Online]. Available:
GraphSniffer model in order to provide higher importance to https://coinmarketcap.com/
[3] W. Chen and Z. Zheng, “Blockchain data analysis: A review of status,
the illicit samples. Table II shows the testing results in term of trends and challenges,” Journal of Computer Research and Development,
Accuracy, Precision, Recall, and F1-Score for the illicit class. vol. 55, no. 9, pp. 1853–1870, 2018.
Accuracy. From the perspective of accuracy, we explore the [4] M. Bartoletti, S. Carta, T. Cimoli, and R. Saia, “Dissecting ponzi
schemes on ethereum: identification, analysis, and impact,” Future
relationship between accuracy and three different models. The Generation Computer Systems, vol. 102, pp. 259–277, 2020.
accuracy experiments results are as follows: [5] M. Vasek and T. Moore, “There’s no free lunch, even using bitcoin:
Tracking the popularity and profits of virtual currency scams,” in
• In Fig. 2, we observe that GraphSniffer outperforms International conference on financial cryptography and data security.
Logistic Regression and Multilayer Perceptron, while Springer, 2015, pp. 44–61.
[6] M. Bartoletti, B. Pes, and S. Serusi, “Data mining for detecting bitcoin
MLP can also achieve a high accuracy rate. Besides, the ponzi schemes,” in 2018 Crypto Valley Conference on Blockchain
results show that the GraphSniffer model is trained well Technology (CVCBT). IEEE, 2018, pp. 75–84.
with only a few epochs. [7] M. Shen, M. Wei, L. Zhu, and M. Wang, “Classification of encrypted
traffic with second-order markov chains and application attribute bi-
• In Fig. 2, the logistic regression does not perform well grams,” IEEE Transactions on Information Forensics and Security,
and show a decrease in accuracy after reaching the peak, vol. 12, no. 8, pp. 1830–1843, 2017.
indicating that it may not suitable for this scenario. [8] K. Henderson, B. Gallagher, T. Eliassi-Rad, H. Tong, S. Basu,
L. Akoglu, D. Koutra, C. Faloutsos, and L. Li, “Rolx: structural role
F1-Score. Since the number of illicit samples is far less extraction & mining in large graphs,” in Proceedings of the 18th ACM
than the number of licit samples, we conduct experiments to SIGKDD international conference on Knowledge discovery and data
mining, 2012, pp. 1231–1239.
obtain the F1-score and present the results as follows: [9] W. Chen, Z. Zheng, J. Cui, E. Ngai, P. Zheng, and Y. Zhou, “Detecting
• In Fig. 3, we can clearly see that the MLP method ponzi schemes on ethereum: Towards healthier blockchain technology,”
in Proceedings of the 2018 world wide web conference, 2018, pp. 1409–
gets better F1-score in previous epochs. But in the later 1418.
training, the GraphSniffer shows higher F1-score than [10] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Federated learning:
other methods, indicating that the graph-based method Challenges, methods, and future directions,” IEEE Signal Processing
Magazine, vol. 37, no. 3, pp. 50–60, 2020.
can take the advantage of graph-structured data. [11] M. Shen, Y. Liu, L. Zhu, X. Du, and J. Hu, “Fine-grained webpage
• On the other hand, due to the input features are quite fingerprinting using only packet length information of encrypted traffic,”
informative already, Multilayer Perceptron has the ability IEEE Transactions on Information Forensics and Security, vol. 16, pp.
2046–2059, 2020.
to train a good model using these features alone. [12] A. Biryukov, D. Khovratovich, and I. Pustogarov, “Deanonymisation
• In Fig. 3, Logistic Regression shows a significant de- of clients in bitcoin p2p network,” in Proceedings of the 2014 ACM
crease in F1-score during the late training period, which SIGSAC Conference on Computer and Communications Security, 2014,
pp. 15–29.
also illustrates that the it cannot solve this problem . [13] F. Tramèr, D. Boneh, and K. Paterson, “Remote side-channel attacks
Summary. Through extensive experimental results, we can on anonymous transactions,” in 29th {USENIX} Security Symposium
({USENIX} Security 20), 2020, pp. 2739–2756.
claim that our GraphSniffer model shows superiority over the [14] M. Shen, J. Zhang, L. Zhu, K. Xu, and X. Du, “Accurate decentralized
existing schemes in terms of model accuracy. application identification via encrypted traffic analysis using graph
neural networks,” IEEE Transactions on Information Forensics and
Security, vol. 16, pp. 2367–2380, 2021.
V. C ONCLUSION [15] S. Delgado-Segura, C. Pérez-Sola, G. Navarro-Arribas, and J. Herrera-
Joancomartı́, “Analysis of the bitcoin utxo set,” in International Confer-
In this paper, we presented a new federated graph neural ence on Financial Cryptography and Data Security. Springer, 2018,
network model GraphSniffer to solve the problem of iden- pp. 78–91.
[16] S. Foley, J. R. Karlsen, and T. J. Putniņš, “Sex, drugs, and bitcoin: How
tifying malicious transactions for digital currency. Different much illegal activity is financed through cryptocurrencies?” The Review
from existing malicious transaction identification schemes, of Financial Studies, vol. 32, no. 5, pp. 1798–1853, 2019.
our approach proposed the use of graph neural network to [17] M. Weber, G. Domeniconi, J. Chen, D. K. I. Weidele, C. Bellei,
T. Robinson, and C. E. Leiserson, “Anti-money laundering in bitcoin:
construct the identification model considering the existence Experimenting with graph convolutional networks for financial foren-
of noteworthy features in the flow of funds between each sics,” arXiv preprint arXiv:1908.02591, 2019.
transaction. To protect the security of the transaction feature [18] D. McGinn, D. Birch, D. Akroyd, M. Molina-Solana, Y. Guo, and
W. J. Knottenbelt, “Visualizing dynamic bitcoin transaction patterns,”
data and the privacy of the model, we introduced federated Big data, vol. 4, no. 2, pp. 109–119, 2016.
learning to realize the joint identification and analysis of [19] W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation
malicious transactions. Extensive experiments conducted on learning on large graphs,” in Proceedings of the 31st International
Conference on Neural Information Processing Systems, 2017, pp. 1025–
the Bitcoin transaction graph dataset demonstrated that our 1035.
GraphSniffer model had excellent capabilities in the task of
malicious transaction identification. In the future, we would

Authorized licensed use limited to: Chinese University of Hong Kong. Downloaded on March 22,2024 at 04:39:36 UTC from IEEE Xplore. Restrictions apply.

You might also like