Leveraging Fuzzy Fingerprints from Large Language Models for Authorship Attribution

Author Attribution has various practical applications such as identifying plagiarism or assigning authorship in historical texts, and plays a crucial role in forensic linguistics. Recent advances have attempted to employ deep learning models for author attribution, resulting in more robust models when compared to previous approaches. In this paper, we introduce a new method to the Author Attribution task and achieve state-of-the-art results in two different datasets. We employ a recent technique that combines Fuzzy Fingerprints with Large Language Models and show that it is possible to obtain unique fingerprints for each author from the datasets. One key feature of this approach is that it allows for a substantial reduction in the size of the output layers of these language models. In addition, we explore the impact of the fingerprint size on the performance of the model and provide illustrative examples of the fingerprints.