Large Language Model XAI approach for illicit activity Investigation in Bitcoin

Large Language Models (LLMs) present an opportunity for interpreting and investigating financial cybercrime activity. Analysts are tasked with learning and understanding the ever-shifting domain of financial crime, coupled with the latest tools being produced to assist them in their investigative tasks. We present an LLM XAI approach for identifying and explaining illicit activity in Bitcoin. We show that LLMs can produce intuitive and highly useful contextual narratives when prompted with Bitcoin transaction data. Extracting embeddings of the narratives allow for the calculation of similarity metrics capable of identifying other illicit transactions. We develop a pipeline of collating similar transactions, creating contextual narratives explaining their similarity, and produce a summary report all for the goal of aiding the investigative analyst.