Abstract
The success of Bitcoin gave rise to the use of blockchain technology. Until recently, blockchains were commonly
deployed as cryptocurrencies. Now, with the rapid development of smart contracts, blockchain technology is finding new
applications. Two main types of blockchain, i.e., public blockchains and permissioned blockchains, were formed.
Nonetheless, a number of issues were overlooked with the fast development of blockchain technology. Existing
blockchain networks and their applications often assume perfect connectivity, leaving the use of blockchains with
connection restriction unaddressed. Moreover, these applications are often deployed on independent networks, resulting
in the formation of isolated data silos. In addition, various financial crimes have also been found on cryptocurrency
networks.
This thesis aims to fill the existing gaps related to the performance, utility improvements and anomaly detection of
blockchain networks. Firstly, we develop mathematical models for transaction processing and block propagation and
conduct an emulation to study the network behaviours under network disturbances. We show node churns do not affect
the transaction throughput, but may cause temporary inconsistencies. We also show block propagation can be modelled
with a modified epidemic-spreading model which incorporates network effects. Secondly, we design and implement two
proxy-based architectures to improve the utility of blockchains. We first propose a blockchain-based architecture to
deliver banking services to regions with an intermittent connectivity to the broader Internet. We next develop a proxybased
framework to enable trusted data transfer between independent permissioned blockchains. Different from existing
interoperability schemes on public blockchains, the proposed framework allows the destination requester to select a set
of peers on the source network to prove the integrity of the transferred data. Thirdly, we investigate the detection of
laundering transactions on the Bitcoin network. We analyse laundering and regular transactions from a graph-theoretic
perspective, and apply manual feature extraction and graph representation learning techniques to build classifiers for
binary classification between laundering and regular transactions. We show that random-walk based embeddings are
effective in binary classification. We then apply the proposed classifiers to predict unknown laundering transactions,