Generation and Classification of Illicit Bitcoin Transactions

Financial fraud is an everyday problem that banking institutions have to face. With the disruption of Bitcoin as a new model which relies on decentralisation and anonymity, attackers have taken advantage of this monetary system. It allows them to obtain funds from illegal activities such as ransomware payments and hide them. At the same time, Law Enforcement Agencies use open-source data to apply network forensics to Blockchain data. The analysis is usually performed by using artificial intelligence. Unfortunately, the current situation shows a scarcity of high-quality data sets to train the detection algorithms. This work tries to overcome this barrier with significant contributions. With nearly 25,000 illicit transactions, we have increased the Elliptic Data Set -the most extensive labelled transaction data publicly available in any cryptocurrency. The former data set only contained 4,545 illicit transactions, resulting in a class imbalance of 9.8:90.2 illicit/licit ratio. Our work has changed that to a 41.2:58.8 illicit/licit ratio. Besides, to show that class imbalance datasets can also be beaten with artificial work, we have studied the use of generative adversarial networks (GAN) for creating synthetic samples. Finally, the last part of this work was dedicated to applying deep learning and, more particularly, long short-term memory networks (LSTM) for the binary classification problem. We show ideal results that can help change the current state-of-the-art trend, mainly focused on machine learning algorithms.