This study focues on mainly fourth methods.

1. How to collect ground truth data for de-anonymization of Bitcoin transactions

2. Generating fine-tuning data and evaluation datasets based on the ground truth data

3. Devising the evaluation metrics to evaluate the performance of Llama2 for de-anonymization of Bitcoin transactions

4. Identifying the best parameters and configurations of Llama2 to de-anonymize Bitcoin transactions

By exploring the dark web and getting Bitcoin transactions data from "ransomwhere" website which provides the Bitcoin transaction information related to ransomware payment, I got ground truth data for this project.

Also, for evaluation, I adopted "Faithfulness" and "BERT-SCORE".

Through these experimens, I found that RAG must be used for de-anonymization of Bitcoin transaction and combining prompt-engineering and adjusting top_similarity_value got higher score for de-anonymization of Bitcoin transactions.

For future works, I plan to improve the accuracy by using more high-quality data from law enforcement agencies.

Also, I aim to explore changing other parameters of RAG and other proposed methods to improve RAG performance.

This research highlights the potential of LLMs to support law enforcement in tracking cybercrime by offering new approach.

By refining these methods, we can take a step towards reducing the crimes exploiting the Bitcoin anonymity.