Digital forensics in law enforcement: A case study of LLM-driven evidence analysis

The advent of digital technology and the ubiquity of mobile devices in today's society has led to a significant increase in the importance of mobile forensics in criminal investigations. Responding to the escalating volume and complexity of data due to enhanced smartphone capabilities and pervasive messaging apps, law enforcement agencies face challenges in data analysis. This study explores improving investigative efficiency through LLM-driven analysis of text from mobile messenger communications. We have conducted experiments on anonymized data collected from real crime scenes by employing three state-of-the-art LLM models, namely GPT-4o, Gemini 1.5 and Claude 3.5. The study focuses on optimizing model performance by employing prompt engineering, interpreting expressions embedded with hidden meanings such as slang, and contextually inferring ambiguous word usage. Finally, model performance is quantitatively evaluated using metrics such as precision, recall, F1 score, and hallucination rate. • LLMs analyzed mobile chat data from real criminal case investigations. • GPT-4o, Gemini 1.5, and Claude 3.5 were benchmarked on forensic tasks. • Prompt engineering improved detection of slang and hidden meanings. • Context-based inference helped resolve ambiguous word usage. • Evaluation included precision, recall, F1 score, and hallucination rate.