This dissertation examines the challenges of detecting illicit activities in cryptocurrency transactions, with a focus on Bitcoin. It begins by analyzing cryptocurrency mixing services and their obfuscation techniques. The research then provides a comprehensive evaluation framework for these services, conducting an assessment of all available services and academic proposals. Following this, the study introduces a novel framework that uses statistical patterns to identify potential money laundering and clustering cryptocurrency addresses that can reveal real-world identities involved in illicit transactions.

The study then leverages the Elliptic dataset, a graph representation of Bitcoin transactions, to classify illicit activities. While classical machine learning methods struggled with the imbalanced nature of financial fraud data, Graph Neural Networks (GNNs) - specifically Graph Convolutional Networks and Graph Attention Networks - proved more effective. By considering the graph topology and connections between nodes, GNNs significantly reduced false negative rates in detecting illicit transactions.

To enhance transparency, the research employs Explainable AI techniques, particularly SHAP values, to interpret the decision-making process of GNN models. This approach not only improves model trustworthiness but also provides insights into the key features and graph structures that contribute to illicit activity detection.

The thesis concludes by presenting a comprehensive toolkit for combating digital financial crimes. It demonstrates that despite the perceived anonymity of blockchain technology, effective methods exist to unveil illicit activities, thus enhancing the security and integrity of cryptocurrency transactions. This work bridges the gap between technological advancement and regulatory compliance, establishing a new standard in the fight against cryptocurrency-based crime.