Detection of illegal money transfers on blockchain networks with graph theory and artificial intelligence methods

Money laundering in blockchain based financial systems poses a significant threat to economic stability because of the anonymity and decentralization of cryptocurrencies. Traditional rule based detection methods generate high false positive rates and struggle to adapt to evolving money laundering strategies. This thesis proposes an innovative model that combines chaotic time series analysis with machine learning and deep learning approaches to detect money laundering activities in blockchain transaction networks. First, the number of structural features extracted from blockchain transaction graphs was increased and transformed into a chaotic time series. The dynamic behavior of these series was characterized by calculating the Lyapunov exponents, and the resulting chaotic features were employed as distinctive indicators of illicit transactions. These features were evaluated through two models. The first model applies machine learning algorithms directly for classification, while the second performs classification using Graph Convolutional Networks on the reconstructed transaction graphs. In addition, a hybrid deep learning model was proposed, combining Convolutional Neural Networks, Bidirectional Long Short Term Memory, and an Attention Mechanism. Experiments conducted on the Elliptic dataset demonstrated that the proposed chaos based feature extraction approach improved the detection accuracy and reduced the false positive rates compared with conventional machine learning and deep learning methods. The findings highlight the strong potential of integrating chaos theory with blockchain transaction analysis to enhance the effectiveness of anti money laundering systems in decentralized finance environments.