ADPP: A NOVEL ANOMALY DETECTION AND PRIVACY-PRESERVING FRAMEWORK USING BLOCKCHAIN AND NEURAL NETWORKS IN TOKENOMICS

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.13, No.6, November 2022
DOI: 10.5121/ijaia.2022.13602 17
ADPP: A NOVEL ANOMALY DETECTION AND
PRIVACY-PRESERVING FRAMEWORK USING
BLOCKCHAIN AND NEURAL NETWORKS IN
TOKENOMICS
Wei Yao, Jingyi Gu, Wenlu Du, Fadi P. Deek and Guiling Wang
New Jersey Institute of Technology, Newark, NJ, USA
ABSTRACT
The increasing popularity of crypto assets has resulted in greater cryptocurrency investor interest and
more exposure in both industry and academia. Despite the substantial socioeconomic benefits, the
anonymous character of cryptocurrency trading makes it prone to abuse and a magnet for illicit purposes,
which cause monetary losses for individual traders and erosion in the standing of the tokenomics industry.
To regulate the illicit behavior and secure users' privacy for cryptocurrency trading, we present an
Anomaly Detection and Privacy-Preserving (ADPP) Framework integrating blockchain and deep learning
technologies. Specifically, ADPP leverages blockchain technologies to build a user management platform
that ensures anonymity and enhances the privacy-preservation of user information. Atop the user
management system, an Anomaly Detection System adapts neural networks and imbalanced learning on
topological cryptocurrency flow among users to identify anomalous addresses and maintain a sanction list
repository. The experiments on the real-world dataset demonstrate the effectiveness and superior
performance of ADPP. The flexible framework can be easily generalized to the crypto assets with public
real-time transaction (e.g., Non-fungible Token), which takes up a significant proportion of market
capitalization in the domain of tokenomics.
KEYWORDS
Cryptocurrency Anomaly Detection, Privacy-Preservation, Graph Neural Networks, Blockchain,
Imbalanced Learning
1. INTRODUCTION
There are over 20,000 cryptocurrencies with over 1 trillion of total market capitalization, as of
June 2022 [1], revealing the trend of technological innovation to digitize in the 21st
century
global economy. The continued evolution and popularity of cryptocurrency is happening for a
reason: it enables efficient payment systems where trading is independent of any political
interference or governmental regulatory bureaucracy, occurring rather through a decentralized
distributed ledger. Yet there exists, for a long time, a discourse regarding the process in which
cryptocurrency transactions can be regulated to prevent criminality. In fact, since its booming, a
widespread illicit behavior in the cryptocurrency markets has been recorded as hackers engage in
phishing attacks and attempt to steal credentials of traders and other sensitive privacy data, such
as account information. These unlawful activities are especially prevalent in the Bitcoin market,
with more than one-quarter of all users and close to one-half of bitcoin transactions being
associated with illegal activities in 2017 and around 76 billion in illegal transactions annually
[2]. Developing countermeasures to mitigate cybercriminality and the illicit use of

18
cryptocurrencies is hence an exceptionally important task to ensure integrity of their transactions
and continued evolution of the digital assets market.
As policymakers focus on regulatory matters, it seems also prudent for researchers and
practitioners to explore technological remedies through user management platforms that can
ultimately enhance the security and preserve the integrity of data and identity as well as to stave
off any potential cybercriminality at the very beginning stages. Nonetheless, user management
itself is a challenging task due to the diversification of end-use applications where one individual
may have multiple virtual identities. Some solutions, like OpenID, can provide more simplicity
allowing a single sign-on function to access different web service providers. In fact, lacking
complete self-management for identities and privacy-preserving of sensitive information still
remains an important open question to be addressed. One approach toward this end has been the
use of traditional Public Key Infrastructure (PKI). However, the PKI may suffer from data loss
and corruption due to the problem of centralized management, such as a Single-Point-of-Failure
(SPOF). Others, like the Liberty Alliance Project [3], advocate for federated authentication
techniques. Such federated methods, although breaking down the centralized management to
several trusted circles, still do not fully resolve the issue when the management systems (i.e.,
service providers) shutdown and become unavailable. Blockchain technology, however, has
proven to be effective in many fields yet has not been sufficiently applied to user management
systems toward mitigating illicit behavior in cryptocurrency trading.
More diverse countermeasures can certainly be coupled with the above-mentioned security
enhanced and privacy-preserving identity management platform to increase effectiveness. As a
primary preventative approach, detection-based countermeasures is a promising solution. These
methods require historical transactions of cryptocurrencies (e.g., Bitcoin) and assume that illicit
transactions happened in the first place. Through the training, the anomaly behavior will be
identified accurately; hence, any future cyber criminality can be prevented. Machine learning
algorithms and more recently neural networks, especially Graph Convolutional Networks (GCN)
[4] and Gated Recurrent Units (GRU), are proposed to train an intelligent detection system that
can evolve over time. These approaches can analyze based on various characteristics, e.g.,
topological networks and temporal interactions, yet they also easily suffer from an imbalanced
class problem. However, findings reveal that they consider only part of these features and are not
capable of learning in a comprehensive view.
Drawing on the above insights from the study, we propose an Anomaly Detection and Privacy
Preserving (ADPP) Framework which fully capitalizes on the advantages of blockchain and deep
learning techniques for cryptocurrency trade. First, leveraging the Self-Sovereign Identity (SSI)
framework that gives individuals control over their own identities, ADPP uses blockchain
technologies to build a user management platform with a workflow of provision, issuance, and
verification. It reduces the loss caused by unidentified users and enhances privacy-preserving of
user identity information protection.
Besides, the Anomaly Detection System (ADS) is smoothly integrated into the platform to
regulate the trade behavior of the crypto market. The nature of ADS is a binary classification
deep model aiming to detect potential illicit addresses. It leverages GCN, GRU, and imbalanced
learning to learn topological and dynamic cryptocurrency transaction networks. The suspicious
illicit addresses detected by ADS are added to the sanction list repository. ADPP will trigger a
warning if illicit addresses are detected during the verification process before two parties start
the trade.
The experiments on a real-world dataset demonstrate the effectiveness of ADPP. The flexible
framework can be generalized to those crypto assets which have public real-time transactions

19
with addresses and features in the market, such as cryptocurrencies and Non-fungible Tokens
(NFTs). Such crypto assets take a significant proportion of market capitalization in the domain of
tokenomics. The main contributions of this paper can be summarized as follows:
1. The blockchain-based privacy-preserving authentication platform leverages the
cryptographic technologies on blockchain ledger to authorize validated users for
cryptocurrency trading without leaking users' privacy.
2. Anomaly Detection System based on GCN, GRU, and imbalanced learning captures
topological patterns and dynamic changes of cryptocurrency flow between users. It aims
to detect potential illicit addresses and maintain a Sanction List Repository to prevent
illicit behaviors.
3. To the best of our knowledge, we are the first to integrate a privacy-preserving
authentication platform with a cryptocurrency anomaly detection model based on deep
learning. The resulting work effectively mitigates cybercriminality by enhancing privacy
and identifying potential illicit addresses. This integration can similarly be generalized to
other crypto assets.
The remaining parts of this paper are organized as follows: Section 2 provides an overview of
related works. Section 3 introduces the technical background. Section 4 presents the architecture
of the proposed platform, detailed system design with core internal workflow, algorithmic
protocol, and a deep learning model-based anomaly detection system. Section 5 presents the
experimental results and critical analysis. Finally, Section 6 offers concluding remarks and
presents research challenges and opportunities to motivate future work.
2. RELATED WORKS
In this section, we introduce related work from three perspectives: identity provider-based
platform, anomaly detection model and any studies that integrate blockchain-based
privacypreserved platform with deep learning technologies. The discussion discloses the dearth
of prior research in privacy-preserving frameworks with functionality of detecting anomaly, and
an urgent need to apply Blockchain technology for protecting user privacy.
2.1. Identity Provider based Platform
Cross-platform and privacy-preserving are two major considerations in terms of designing a
identity management system. Aiming toward cross-platform identity management, Microsoft
experimented with Identity Management (IdM) on Microsoft.NET Passport [5], allowing users to
log in to multiple service providers' websites with the same identity provided by Microsoft as the
Identity Provider (IdP). The IdP allows users to access the different websites and redirects to
different user profiles through the identity managed by the IdP. However, the drawbacks of
centralized management are apparent, as all service providers and customers are affected if the
Identity Provider (IdP) becomes unavailable or as a result of data loss and corruption. To tackle
the problem of centralization, the Liberty Alliance Project [3] made another attempt at the IdP to
build a federated network identity management. It advocates for a federated IdM method, which
focuses on establishing trust circles between service providers and federated isolated customers.
However, the problem with federated IdM is that once an authenticated service provider is
unavailable, customers will not be able to access the resources of that provider. From the security
perspective, OAuth (Open Authorization) [6] allows service providers' websites to safely access
various identity information stored in the IdP with the user's authorization. In the authentication
process, the customer can access the service provider's website without providing their account
password but in the form of a token to the service provider's website. OAuth is widely used in

20
the current internet environment since its usability and performance are improved. However,
OAuth still suffers from some security vulnerabilities [7]. Another disadvantage of the OAuth
framework is that the IdP serves as a centralized identity provider, which in some situations may
yield data loss. For instance, if when accessing any website using OAuth provided by Google
and it blocks that service provider, the customer will lose the access rights and data on that
website.
2.2. Anomaly Detection
Artificial Intelligence has been increasingly playing an essential role in detecting and preventing
cybercrime [8]. Standard machine learning algorithms (e.g., Random Forest [9]) can be applied
to anomaly detection tasks which, by its nature, are binary classification problems. Such
algorithms are demonstrated to be effective [10, 11]. These methods can extract phishing fraud
features from the data collected and show a stable performance, however, they do not leverage
any graph information. Since cryptocurrency trading can be naturally formulated as a transaction
graph, a line of work [12, 13, 14] employed graph-based algorithms to detect illicit behavior.
With the continued growth of deep learning, graph neural networks, such as GCN [4] and graph
attention networks (GAT) [15] have been applied to detect malicious transactions. Based on an
analysis of the weaknesses of attackers a heterogeneous graph neural network with an attention
mechanism is also proposed [13]. The main drawback of the graph-based algorithms is the
neglect of the temporal dependencies of multivariate time series data as identifying an anomaly
in real time is a known challenge due to the frequency of seasonal behavior or a change in trend
[16]. Statistical models, such as Autoregressive Integrated Moving Average (ARIMA) [17], and
recurrent neural networks, such as Long Short-Term Memory (LSTM) [18] and GRU [19] are
common approaches to address the temporal dependencies. Another challenge is the imbalance
problem in the transaction dataset where illicit transactions are minority and only occupy a small
portion of the whole. Some existing work leverages the imbalanced learning techniques:
sampling methods drop majority samples or repeat minority samples randomly to control the
imbalance ratio directly [20]; and re-weighting algorithms adaptively assign and adjust the
weight on samples based on their classification performance [21]. The advantage of our work
over existing approaches is that we consider both temporal dependence and the imbalance
problem in anomaly detection.
2.3. Blockchain-Based Privacy-Preserved Platform Integrated with Machine
Learning (ML)/Deep Learning (DL)
There exists some efforts toward integrating blockchain with ML/DL techniques. A privacy
preserving framework in smart power networks has been presented [22]. Specifically, a
blockchain-based two-level privacy module is designed to verify data integrity and LSTM is
employed for anomaly detection. A similar idea is proposed [23], where a two-level privacy
preservation approach with a blockchain module to transmit data and a standard machine
learning algorithm (i.e., Principal Component Analysis) for data transformation is designed.
Such integration of blockchain with the ML technique is applied into Internet of Things (IoT)-
driven smart cities. However, we found no work that integrates blockchain with ML/DL in
cryptocurrency trading. Additionally, the integration frameworks proposed are generally tailored
for specific domains and thus difficult to migrate to other fields.

21
3. TECHNICAL BACKGROUND
In this section, the basic underlying knowledge regarding privacy in cryptocurrency trading and
blockchain-based identity management framework will be described. In addition, the deep
learning algorithms that we build upon are described.
3.1. Privacy Preserving
Privacy preserving involves two aspects. (1) Transaction Privacy refers to the identity
information in transactions of cryptocurrency trading, which is essential since the transactions
are published on a public blockchain. To achieve anonymity and protect transaction privacy,
cryptocurrency platforms, such as Bitcoin, use the address as the presentation for the trading
parties, usually called pseudonymous, to hide the real identity information of both parties. The
addresses are essentially the hash values of the public keys of both parties. (2) Identity Privacy
refers to the identity information of the two parties of a cryptocurrency transaction in the real
world. Although the user's identity in a cryptocurrency transaction is not associated with the
identity in the real world, there are still some situations that may result in identity privacy
leaking. For instance, by leveraging clustering and other technologies to analyze a transaction,
find out the transaction relationship map between different accounts, and infer the input and
output of the transaction, the real identity information of the two parties can finally be targeted.
Since the blockchains for cryptocurrency trading use pseudonym addresses to protect transaction
privacy, the framework proposed in this paper mainly focuses on identity privacy.
3.2. Self-Sovereign Identity
Self-Sovereign Identity (SSI) is a novel distributed identity management framework that gives
individuals control over their digital and physical identities and is mostly built upon a
blockchain. Traditionally users manage their digital identities either through their accounts on
each identity provider, or by relying on a manager of different identity providers such as Google
Sign-In. However, traditional methodologies result in (a) neither identity providers allowing
users to control their information themselves; and (b) a limitation if users want to present their
identity information to other individuals or service providers. In the latter case, the identity
providers, who have full control, must provide approaches for the verification process. In other
words, the users cannot control the information presented. In an SSI system, contrarily, if a
particular identity provider authorizes the identities, these are maintained by the individuals, not
the provider. To achieve individual control over identity information, SSI employs blockchain
technology: the identity providers can publish their cryptographic authentication on the
Distributed Ledger Technology (DLT); thus, other individuals can cryptographically verify the
identities without interaction with the identity providers.
There are two major standard specifications. (1) Decentralized Identifiers (DID) is a W3C
standard specification [24] for SSI. A DID can be resolved to a DID Document that provides the
subject's identifying information (to whom the DID belongs). The DID Document contains the
subject's authentication information, such as the public keys needed to validate the subject's
signature. It also includes service endpoints that identify the URL where the subject's verification
information can be retrieved. (2) Verifiable Credential (VC) is a W3C standard specification
[25] for DID-based cryptographically verifiable digital credentials. It can represent the same
information as a physical credential. With blockchain technologies, a VC can be issued by an
issuer in a more tamper-evident and trustworthy method to a holder. A VC contains claims about
the holder, certified by the issuer.

22
In addition, Wallet is an application that generates private and public key pairs and securely
stores the key-pairs. The private keys can be used to prove ownership of DIDs and VCs
cryptographically. The received credentials are also stored in the wallet.
3.3. Graph Convolutional Network
Traditional convolutional networks capture local information by a sliding filter on images or
grids. GCN is the generalization form of convolutional network on graphs, which extracts the
topological features to generate new node representations. It automatically learns not only the
characteristics of central nodes but also the associated information from connected neighbors.
Given an undirected graph 𝐺 = (𝑉, 𝐸) with 𝑁 = |𝑉| nodes and 𝑀 = |𝐸| edges, 𝐴 ∈ ℝ𝑁 × 𝑁 is the
adjacency matrix of 𝐺 where each element 𝐴𝑖,𝑗 is 1 if node 𝑖 and 𝑗 is connected, 0 otherwise. 𝑋 ∈
ℝ𝑁 × 𝑃 is a feature matrix. GCN is denoted as:
where 𝐴 is the adjacency matrix normalized by the neighbors' degree. The self-loop constant
matrix 𝐼 considers the influence of the node itself. The degree matrix 𝐷 ∈ ℝ𝑁 × 𝑁 is a diagonal
matrix composed of node degrees. 𝑊 is a trainable weight matrix and σ(⋅) represents the
activation function. We refer the reader to [4] for more details.
3.4. Gated Recurrent Units
GRU is simple, yet effective and fast to obtain time series patterns. Let 𝑋𝑡 denote the training
input. The key design of GRU consists of reset gate 𝑅𝑡 takes account of short-term memory,
determining how much the previous states are kept. Update gate 𝑍𝑡 is responsible for long-term
memory, controlling how much new information to pass along into the future.
𝑅𝑡= σ(𝑋𝑡𝑊𝑥𝑟+ 𝐻𝑡−1𝑊ℎ𝑟+ 𝑏𝑟) (2)
𝑍𝑡= σ(𝑋𝑡𝑊𝑥𝑧+ 𝐻𝑡−1𝑊ℎ𝑧+ 𝑏𝑧) (3)
𝐻𝑡= 𝑡𝑎𝑛ℎ(𝑋𝑡𝑊𝑥ℎ+ (𝑅𝑡⊙ 𝐻𝑡−1)𝑊ℎℎ+ 𝑏ℎ) (4)
𝐻𝑡= 𝑍𝑡⊙ 𝐻𝑡−1 + (1 − 𝑍𝑡) ⊙ 𝐻𝑡 (5)
where 𝐻𝑡−1 is the hidden state at 𝑡 − 1, 𝑊𝑥𝑟, 𝑊ℎ𝑟, 𝑊𝑥𝑧 and 𝑊ℎ𝑧 are weight parameters, 𝑏𝑟 and 𝑏𝑧
are the bias parameters. Sigmoid function σ(⋅) maps the value to the interval of (0,1).
To incorporate the effect of reset gate 𝑅𝑡, candidate hidden state 𝐻𝑡 is constrained by the
activation function in the interval (−1,1). The element-wise product operator ⊙ controls which
matrix to remove from the previous time step. If entries in 𝑅𝑡 are close to 0, then it ignores 𝐻𝑡−1
and focuses on 𝑋𝑡 only. Finally, the update gate 𝑍𝑡 determines how much information from 𝐻𝑡−1
and 𝐻𝑡 to be included in 𝐻𝑡. Suppose the update gates in all time steps are 1, then the hidden state
of the beginning will be kept and passed to the final output, no matter how long the period is. We
direct the readers to [18] if interested.

23
4. ADPP
4.1. The Architecture of ADPP
Figure 1. ADPP Architecture
Figure 1 presents the architecture of ADPP. ADPP defines three entities. (1) Trusted
Committees (TCs) have the highest authority. They are pre-selected and pre-configured in the
blockchain and host the other two consortium DLTs. An example TC can be the U.S. Securities
and Exchange Commission (SEC). (2) Trading Providers (TPs) are authorized by a TC. An
example TP can be Coinbase. It obtains licence from SEC and issues credentials to its customers.
(3) Peer customers (PCs) are users who trade cryptocurrencies with each other in the ADPP
after authorized by a TP.
ADPP has five components. (1) Anomaly Detection System (ADS) is responsible for detecting
suspicious anomalous addresses. (2) Sanction List Repository (SLR) maintains addresses of

24
anomalous customers that are caught by ADS. (3) Peer Customer Distributed Ledger
Technology (PC-DLT) stores transactions that record the credentials issued by TPs to the PCs.
(4) Trading Providers Distributed Ledger Technology (TP-DLT) is used to store transactions
that a TC authorizes a TP. (5) Virtual Broker (VB) is a suite of middleware. It can be a web
application containing some subsystems that leverage RESTful API and Web Service to provide
the entities with certain features, such as provision, issuance, verification, and key management.
The workflow of ADPP is shown in Figure 1. TCs first complete their initialization and
provision process. TCs can then evaluate and authorize TPs who apply to establish a trading
platform. Once a TP is authorized, a corresponding blockchain transaction is published on the
TP-DLT. Authorized TPs then issue VCs to their customer PCs through blockchain transactions
on the PCDLT. With more transaction data available, ADS learns anomalous behaviors from
crypto transaction data. The suspicious illicit addresses detected by ADS are added into SLR.
When two PCs want to trade cryptocurrency, they can verify each other's credentials and check
whether the counterparty is in the SLR. Note that ADS can either be maintained by TC or the TP
consortium.
4.2. Privacy-Preserving Authentication Platform
As shown in Figure 2, our blockchain-based privacy-preserving authentication platform allows
users to regain control of identities and personal data to a certain extent. A decentralized key
management system is integrated into the system to allow trading institutions to issue credentials
to valid users. Only authenticated users are allowed cryptocurrency trading.
4.2.1. Provision
(1) TC Provision: All TC committee members are pre-selected, and the verifiable cryptographic
information is stored in the genesis block of the TP-DLT. (2) TP Provision: A TP can use the
tools provided by the TC to generate a key pair and a DID. The DID is linked to the key pair by
using the public key as a unique identifier in the DID. The registration of the TP’s DID with the
TC is completed off-chain by submitting the public key and other information directly to the TC.
When the off-chain authorization is completed, the TC writes a transaction into the TP-DLT,
indicating that the TP is authorized and can issue credentials to PCs. Then the TP creates and
registers a credential schema in the PC-DLT. The credential schema describes the set of
attributes for a particular certificate. The schema defines what information a PC requires to trade
in the TP's platform. Based on the credential schema, the TP creates and registers a credential
definition in the PC-DLT. The TP's identity information and the cryptographic attributes is
bound to the credential schema and will be used to verify the credential. (3) PC Provision: A PC
downloads a wallet and generates a key-pair and a DID. The key-pair is stored on the mobile
device and must only be accessed by the PC’s password or biometric authentication.

25
ADPP Proposed Protocol Description
Figure 2. Protocol Description
4.2.2. Credential Issuance
A PC obtains a verifiable credential from a TP. Verifiable credentials must contain the PC’s
information affirming the PC is legal to perform a particular cryptocurrency trading. If a
Dashed arrows: a value from Ledger
Green: interaction with wallet
Blue: interaction with PC-DLT Ledger
Black: communication between
Issuer/Holder/Verifier

26
credential schema for a particular kind of cryptocurrency has been registered on the PC-DLT and
the PC wants to acquire a credential for trading that cryptocurrency, the PC will send a credential
schema request to the PC-DLT to use the existing schema. Otherwise, the PC must contact the
TP about registering a new schema. Once the PC receives the credential schema from PC-DLT,
it can provide its DID and the value of required attributes in the schema and send a credential
request to the TP. The validation process for the PC's information in the credential request is
excluded from the scope. The TP ensures that the information provided by the PC is correct and
places the
4.2.3. Credential Verification
Before two PCs trade cryptocurrency, mutual authentication is required. In a one-way
authentication, the verifier PC sends a proof request for particular attributes to the holder PC. For
example, in the case of Bitcoin trading, the proof request must contain the address attribute for
Bitcoin transactions, along with a valid credential issued by a TP. After the holder PC receives
the proof request from the verifier PC, it checks whether it has a corresponding schema
containing the requested attributes. If yes, the holder PC can generate a presentation of the
credential, which contains the requested attributes, and sends the presentation to the verifier PC.
The verifier PC first checks whether a trustworthy TP has issued the credential contained in the
presentation. Then it verifies whether the presented attributes in the proof suffice the request. In
the Bitcoin case, the verifier will check whether a TP authorizes the holder’s address for a
Bitcoin transaction. Both parties will also check whether the counterparty is listed in the SLR.
4.3. Anomaly Detection System
The anonymous character of cryptocurrency trading is often abused for illicit purposes, such as
ransomware, scam, terrorism, etc. PCs that initiate or receive cryptocurrency transactions with
such illegal behaviors are categorized as anomalous users. To prevent malicious usage of
cryptocurrency, ADS is designed as a binary classification model that formulates cryptocurrency
transactions into graphs, aiming to identify anomalous addresses and add them to the SLR.
4.3.1. Graph Formulation
Information regarding cryptocurrency transactions between addresses is publicly available. Each
transaction is associated with a sender address, a receiver address, a time stamp, and a number of
features. The time stamp from raw data indicates the specific time when the cryptocurrency
blockchain confirms the transaction. The confirmation time varies from minutes to hours,
depending on the network. A time step in our model is defined as an interval of multiple days,
which includes multiple time stamps so that various transactions in this interval can be
graphically aggregated. The transaction features, such as transaction volume, cryptocurrency
value, and transaction fee, are shared by both addresses.
At each time step 𝑡, cryptocurrency transactions in this step are used to construct an undirected
graph 𝐺𝑡 = (𝑉𝑡,𝐸𝑡) with adjacency matrix 𝐴𝑡 and node features 𝑋𝑡. Each edge (𝑣𝑡,𝑎, 𝑣𝑡,𝑏)
represents the flow of cryptocurrency in a specific transaction, the nodes 𝑣𝑡,𝑎 and 𝑣𝑡,𝑏 represent
the sender address and receiver address in this transaction, respectively. If there are multiple
transactions between two addresses in the time step 𝑡, the edges and the nodes can be
differentiated by transaction id. There are two types of node features 𝑋𝑡: transaction features
shared by the same edge, and the aggregated features derived from neighbors of node in the
transaction network (e.g., statistics of neighboring transactions, the length of the transaction
chain, etc.)

27
Therefore, given a set of graphs in the historical 𝑇 time periods, we can learn an anomaly
detection model 𝐹 to classify whether each node in the graph at time step 𝑡 is problematic or not.
𝑌̂𝑡= 𝐹(𝐺𝑡−𝑇+1,… , 𝐺𝑡) is a set of predictions for the nodes 𝑉𝑡. Each element 𝑌̂𝑣,𝑡 ∈ {0,1} is a binary
value, 1 if 𝑣𝑡 is anomalous, 0 otherwise.
4.3.2. Anomaly Detection Model
To extract topological dependence and temporal dynamics simultaneously from transaction
networks, ADS integrates GCN with GRU, which are introduced in the previous section, into a
sequence of layers, as shown in Figure 3. For a single layer from time step 𝑡 to 𝑡 + 1, both
feature matrix 𝑋𝑡 and adjacency matrix 𝐴𝑡 are first fed into GCN; then the output 𝑋𝑡′ becomes the
input of the GRU. 𝐻𝑡 is the output of the current layer, and is fed as the input to the next layer. At
the last time step, a function 𝑓 concatenating Multilayer Perceptrons (MLPs) and activation
functions maps 𝐻𝑡 to the interval (0,1). Each row of the output 𝐻𝑡′ 𝑁×2 contains two
values representing the probability of each node belonging to two classes. The final prediction 𝑌̂𝑡
𝑅𝑁 denotes the class with higher probability.
The node classified by the model represents the address in a specific transaction. An address
may appear in multiple nodes which corresponds to different transactions. If an address is
detected as anomalous in any node, it is treated as an anomalous address and will be added into
the SLR, considering an illicit address can also conduct legal transaction but a one-time good
behavior won't make it a licit address.
Figure 3. Overall Structure of Anomaly Detection Model
4.3.3. Enhancements Toward Imbalanced Class Problem
The anomaly detection task often suffers from an imbalanced class problem which shows a
highly skewed class distribution. In our task, the proportion of malicious addresses is far less
than the licit ones. To deal with this issue, ADS implements two enhancements in the proposed
networks. For the simplicity of notation, the time step 𝑡 is omitted.
Sampling Procedure is designed to balance the class of samples. Before training the networks,
a subset of nodes is randomly sampled to construct the subgraph. The sampling probability for

28
each node is determined by two aspects. (1) In bitcoin, the malicious node is the minority class,
and the licit node is the majority class. The minority class is preferred to be chosen in the
sampling procedure. (2) In the graph networks, the node with a larger degree is considered much
more important. Such a node deserves to have a higher probability. Hence, the sampling
probability 𝑃𝑣
where 𝐷𝑣 is the degree of node 𝑣, 𝑐𝑣 is the total number of samples in the class that node 𝑣
belongs to. The sampling probability is normalized so that the sum equals to 1. We randomly
choose nodes with corresponding sampling probability by time steps to generate a set of
subnodes and construct the subgraphs. ADS is trained by the subnodes.
Cost-Sensitive Loss Function is employed to reduce the influence of the imbalanced class. It
assigns weights to classes during the training process. We employ cross entropy in this binary
classification task to optimize the model. Let 𝐶1, 𝐶2 denote the majority (licit) and minority
(anomalous) classes respectively, |𝐶1| and |𝐶2| denote the total samples in each class, the
weights α1 and 𝛼2 are assigned to each class. The cost of missing an anomalous address is much
larger than recognizing licit nodes as anomalous one. Hence, the weight for malicious samples
𝛼2 is set to be larger than 𝛼 in the weighted entropy loss:
where 𝑌̂𝑣 is the true label of node 𝑣, 𝐻𝑣′ is the probability of node 𝑣 being anomalous.
Besides weighted entropy loss, L2 regularization ℒℒ2 is added to the loss function. It uses weights
of GCN in all historical time steps as the constraint of model complexity. The L2 regularization
helps prevent overfitting during the training process. 𝜆1 and 𝜆2 are penalty parameters to balance
two types of loss.
5. EXPERIMENTS
Robust experiments were designed and carried out over real-world dataset to demonstrate the
effectiveness of our proposed framework ADPP.
5.1. Datasets
Elliptic Data Set is released by Elliptic, a blockchain and cryptocurrency analytics company. It is
the largest labeled cryptocurrency transaction data publicly available
(https://www.kaggle.com/datasets/ellipticco/elliptic-data-set). The dataset is a bitcoin transaction
network containing 203,769 nodes and 234,355 edges over 50-time steps. Each time step denotes
an interval of 14 days. Each node is associated with 166 features, all of which are derived from
public information and a label indicating whether it belongs to an illicit or licit class. The overall
imbalance ratio of majority over minority class is approximately 10:1. Before the training

29
process, the size of the dataset is shrunk to eighth of its original size by the sampling procedure.
To tune the parameters and evaluate the performance, the dataset is split into training, validation,
and test sets along time steps as the ratio of 31:5:13.
5.2. Implementation Setting
The two distributed ledgers in ADPP are deployed by leveraging Hyperledger Indy, an
opensource project for consortium blockchain [26] and is funded by Linux Foundation. To fit the
minimum number of nodes for a consensus model that can solve the Byzantine Fault Tolerance
(BFT) problem [27]. The PC-DLT and TP-DLT are deployed on four TC nodes.
ADS is implemented in PyTorch. Hyper-parameters are tuned based on the validation set of
Elliptic Data Set. We use Adam [28] as the optimizer with 0.001 of learning rate, 800 epochs.
The historical length of transaction graph 𝑇 is set to 5. The dimension of GCN and GRU layers
are set to 128. The penalty parameters 𝜆1, 𝜆2, and weighted entropy parameters 𝛼1, 𝛼2 in the cost
sensitive loss function are set to 1, 0.001, 0.35 and 0.65, respectively.
5.3. Performance Analysis
We implement ADPP on Elliptic Data Set to evaluate performance regarding precision, recall,
and F1. Precision evaluates the percentage of nodes classified as illicit are indeed illicit. Recall
measures the percentage of illicit nodes are correctly identified. F1 is a combination of two
metrics.
Figure 4 (a) presents the classification performance and the proportion of illicit samples over
time. In the first six-time steps of the testing period, the performance of ADPP is steady at
around 70% of precision, 50% of recall and 55% of F1 on average. It reveals that half of illicit
addresses in the transactions are caught by the detection model. Note that the task is challenging
given that the data is very imbalanced and only 10% of samples are illicit. Over time, the model
achieves 80% of all metrics at 43rd
step, demonstrating our model's effectiveness. Note that from
the 44th
to the 47th
step, the percentage of anomalous samples sharply decreases to almost 0,
probably due to the shutdown of some dark markets. Correspondingly, the performance of our
model decreases. However, ADS can still catch half of the illicit nodes at the cost of
misclassifying some good nodes as problematic. We argue that it is much more important to
detect anomalous addresses even at the expense of recognizing legal addresses as illicit ones,
considering the latter case is less harmful.
To quantify the efficacy of the enhancements towards imbalanced class problems, two variants
of ADPP are implemented on the same test period of the Elliptic Data Set. Figure 4 (b) shows
the variant without a sampling procedure in the training process. At the first period when there
exists a dark market and a large number of anomalous addresses and transactions, 20% of them
are missed by the variant compared with ADPP. During the shutdown of the dark market, the
variant is not able to detect negative nodes. Compared with the variant in Figure 4 (c), which
assigns equal weights on both classes in the loss function, i.e., 𝛼1 = 𝛼2, the cost-sensitive
learning enables ADPP to focus much on minority samples. It largely boosts the performance
regarding Precision and F1. This demonstrates that our proposed ADPP framework is effective
for extracting illicit patterns and identify anomalous addresses.

30
Figure 4. Anomaly detection performance results in testing periods over temporal dimension.
The left subfigure (a) is the performance of ADPP. The middle subfigure (b) is the performance of variant
of ADPP without sampling. The right subfigure (c) is the performance of variant without cost-sensitive
learning.
6. CONCLUSION
This paper presents an anomaly detection and privacy-preserving platform (ADPP) integrating
blockchain and deep learning techniques for crypto market trade. This blockchain-based privacy
preserving authentication platform enhances the security and privacy of user information. The
Anomaly Detection System identifies illicit addresses in the transactions. The detected illicit
users/addresses are added to the Sanction List Repository to help regulate the behavior in the
crypto market.
The experiments are conducted on real-world cryptocurrency transaction data. Over half of the
illicit addresses in the transactions can be detected by ADS, whether it’s under the dark market
or not. ADPP will trigger a warning if anomalous users in SLR are involved in a trade. It
demonstrates the effectiveness of ADPP. It has a high potential for generalization ability to
various crypto assets.
7. LIMITATIONS AND FUTURE WORKS
In the future, our work can be enhanced from three perspectives: (1) Our current work focuses
on trade in the cryptocurrency market only. The improvement towards a universal framework for
all types of crypto assets is a promising direction. (2) The performance of Create, Read, Update,
and Delete (CRUD) operations in SLR can be further enhanced. For instance, the complexity of
the searching operation can be reduced by adopting a probabilistic data structure such as bloom
filter. (3) The training of ADS currently relies on external data. From our existing DLTs,
additional data can be collected and applied to the federated learning framework to make it more
robust and secure.
ACKNOWLEDGEMENTS
The research is partially supported by FHWA EAR 693JJ320C000021.
REFERENCES
[1] E. A. Akartuna, S. D. Johnson, and A. Thornton, “Preventing the money laundering and terrorist
financing risks of emerging technologies: An international policy delphi study,” Technological
Forecasting and Social Change, vol. 179, p. 121632, 2022.

31
[2] S. Foley, J. R. Karlsen, and T. J. Putnin, š, “Sex, drugs, and bitcoin: How much illegal activity is
financed through cryptocurrencies?,” The Review of Financial Studies, vol. 32, no. 5, pp. 1798– 1853,
2019.
[3] S. Cantor, J. Hodges, J. Kemp, and P. Thompson, “Liberty id-ff architecture overview,” Wason,
Thomas (Herausgeber): Liberty Alliance Project Version, vol. 1, 2003.
[4] D. K. Duvenaud, D. Maclaurin, J. Iparraguirre, R. Bombarell, T. Hirzel, A. Aspuru-Guzik, and R. P.
Adams, “Convolutional networks on graphs for learning molecular fingerprints,” Advances in neural
information processing systems, vol. 28, 2015.
[5] R. Oppliger, “Microsoft. net passport: A security analysis,” Computer, vol. 36, no. 7, pp. 29–35,
2003.
[6] B. Leiba, “Oauth web authorization protocol,” IEEE Internet Computing, vol. 16, no. 1, pp. 74– 77,
2012.
[7] S. Simpson and T. Groß, “A survey of security analysis in federated identity management,” in IFIP
International Summer School on Privacy and Identity Management, pp. 231–247, Springer, 2016.
[8] S. Dilek, H. Cakır, and M. Aydın, “Applications of Artificial Intelligence Techniques to Combating
Cyber Crimes: A Review,” IJAIA, vol. 6, no. 1, pp. 21–39, Jan. 2015, doi: 10.5121/ijaia.2015.6102.
[9] G. Biau and E. Scornet, “A random forest guided tour,” Test, vol. 25, no. 2, pp. 197–227, 2016.
[10] E. Rabieinejad, A. Yazdinejad, and R. M. Parizi, “A deep learning model for threat hunting in
ethereum blockchain,” in 2021 IEEE 20th International Conference on Trust, Security and Privacy in
Computing and Communications, pp. 1185–1190, IEEE, 2021.
[11] M. Weber, G. Domeniconi, J. Chen, D. K. I. Weidele, C. Bellei, T. Robinson, and C. E. Leiserson,
“Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial
forensics,” arXiv preprint arXiv:1908.02591, 2019.
[12] X. Ao, Y. Liu, Z. Qin, Y. Sun, and Q. He, “Temporal high-order proximity aware behavior analysis
on ethereum,” World Wide Web, vol. 24, no. 5, pp. 1565–1585, 2021.
[13] Z. Liu, C. Chen, X. Yang, J. Zhou, X. Li, and L. Song, “Heterogeneous graph neural networks for
malicious account detection,” in Proceedings of the 27th ACM International Conference on
Information and Knowledge Management, pp. 2077–2085, 2018.
[14] N. Pocher, M. Zichichi, F. Merizzi, M. Z. Shafiq, and S. Ferretti, “Detecting anomalous
cryptocurrency transactions: an aml/cft application of machine learning-based forensics,” arXiv
preprint arXiv:2206.04803, 2022.
[15] P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y. Bengio, “Graph attention
networks,” arXiv preprint arXiv:1710.10903, 2017.
[16] S. Ahmad, A. Lavin, S. Purdy, and Z. Agha, “Unsupervised real-time anomaly detection for
streaming data,” Neurocomputing, vol. 262, pp. 134–147, 2017.
[17] J. D. Hamilton, Time series analysis. Princeton university press, 2020.
[18] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp.
1735–1780, 1997.
[19] K. Cho, B. Van Merriënboer, D. Bahdanau, and Y. Bengio, “On the properties of neural machine
translation: Encoder-decoder approaches,” arXiv preprint arXiv:1409.1259, 2014.
[20] Y. Liu, X. Ao, Z. Qin, J. Chi, J. Feng, H. Yang, and Q. He, “Pick and choose: a gnn-based
imbalanced learning approach for fraud detection,” in Proceedings of the Web Conference 2021, pp.
3168–3177, 2021.
[21] X. Li, X. Sun, Y. Meng, J. Liang, F. Wu, and J. Li, “Dice loss for data-imbalanced nlp tasks,” arXiv
preprint arXiv:1911.02855, 2019.
[22] M. Keshk, B. Turnbull, N. Moustafa, D. Vatsalan, and K.-K. R. Choo, “A privacy-
preservingframework-based blockchain and deep learning for protecting smart power networks,”
IEEE Transactions on Industrial Informatics, vol. 16, no. 8, pp. 5110–5118, 2019.
[23] P. Kumar, R. Kumar, G. Srivastava, G. P. Gupta, R. Tripathi, T. R. Gadekallu, and N. N. Xiong,
“Ppsf: a privacy-preserving and secure framework using blockchain-based machinelearning for iot-
driven smart cities,” IEEE Transactions on Network Science and Engineering, vol. 8, no. 3, pp.
2326–2341, 2021.
[24] M. Sporny, D. Longley, M. Sabadello, D. Reed, O. Steele, and C. Allen, “Decentralized identifiers
(dids) v1.0.”
[25] M. Sporny, D. Longley, and D. Chadwick, “Verifiable credentials data model v2.0.”
[26] W. Yao, J. Ye, R. Murimi, and G. Wang, “A survey on consortium blockchain consensus
mechanisms,” arXiv preprint arXiv:2102.12058, 2021.

32
[27] L. Lamport, R. Shostak, and M. Pease, “The Byzantine Generals Problem,” ACM Transactions on
Programming Languages and Systems, vol. 4, no. 3, p. 20, 1982.
[28] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint
arXiv:1412.6980, 2014.
AUTHORS
Wei Yao received the B.S. degree in Computer Science from Huazhong University
of Science and Technology, Wuhan, China, M.S. degree in Electronic Engineering
from University of Hartford, and M.S. degree in Computer Science in Central
Connecticut State University. Wei is currently a Ph.D. candidate in Computer
Science at New Jersey Institute of Technology. His research interests include Internet
of Things, cybersecurity, blockchain technologies and applications in various areas.
Jingyi Gu received the B.S. degree in Economic Statistics from Xiamen University in
2017, and the M.S. degree in Data Science (Statistics track) in Rutgers University,
New Brunswick, in 2019. She is currently a Ph.D. candidate in Computer Science at
New Jersey Institute of Technology. Her research interests include deep learning and
data mining applications in various areas.
Wenlu Du received the B.S. degree in Software Engineering from Dalian Jiaotong
University in 2013, and the M.S. degree in Computer Science in Worcester
Polytechnic Institute in 2015. She is currently a Ph.D. candidate in Computer Science
at New Jersey Institute of Technology. Her research interests include reinforcement
learning and deep learning applications.
Dr. Fadi P. Deek is a Distinguished Professor at New Jersey Institute of Technology
(NJIT) where he began his academic career as a student in the early 1980s. He
received his B.S. and M.S. in Computer Science, and PhD in Computer and
Information Science; all from NJIT. Dr. Deek’s faculty appointments are in two
departments: Informatics (in the College of Computing) and Mathematical Sciences
(in the College of Science and Liberal Arts). Dr. Deek also serves as a member of the
Graduate Faculty at Rutgers University-Newark.
Dr. Guiling (Grace) Wang is currently a Distinguished Professor and Associate Dean
for Research at Ying Wu College of Computing at the New Jersey Institute of
Technology (NJIT). With a primary appointment at the Computer Science Department,
she has held a joint appointment at the MT School of Management and at the
Department of Data Science. Dr. Wang received a B.S. in Software from Nankai
University. She received her Ph.D. in Computer Science and Engineering and a minor
in Statistics from The Pennsylvania State University in May 2006 and joined NJIT as
an Assistant Professor the same year. In 2019, she obtained the Chartered Financial Analyst (CFA)
designation. She established NJIT’s AI Center for Research in 2021 and has become the Founding Director
since then. In 2022, she was elevated to IEEE Fellow, becoming the first female IEEE Fellow at NJIT. Her
research interests include FinTech, deep learning, blockchain technologies, and intelligent transportation.

ADPP: A NOVEL ANOMALY DETECTION AND PRIVACY-PRESERVING FRAMEWORK USING BLOCKCHAIN AND NEURAL NETWORKS IN TOKENOMICS

More Related Content

Similar to ADPP: A NOVEL ANOMALY DETECTION AND PRIVACY-PRESERVING FRAMEWORK USING BLOCKCHAIN AND NEURAL NETWORKS IN TOKENOMICS (20)

Recently uploaded (20)

ADPP: A NOVEL ANOMALY DETECTION AND PRIVACY-PRESERVING FRAMEWORK USING BLOCKCHAIN AND NEURAL NETWORKS IN TOKENOMICS