Year: 2025
Type: dissertation
Abstract: According to GDPR articles 7, 17, and 25, companies are required by law to be transparent with personal information stored in their database. Organizations that allow users to save unstructured data struggle to comply with GDPR requirements. Especially when personal information is saved and needs to be managed. This becomes a time consuming task when managing larger sets of data. This study evaluated LLMs in alocal environment to help system administrators comply with GDPR requirements on Swedish unstructured data from Metria’s Markkoll containing personal information. Previous studies utilize powerful hardware and mixed technologies. This study focused on smaller open-source LLMs in a local environment to provide a cost-effective alternative. To answer the specified problems, the study conducted a controlled experiment to benchmark models straight out of the box and a comparative user evaluation to test the system’s capability to help users comply with GDPR requirements. The two best Gemma3:12B and Llama3.1:8B scored a F1-score of 0.722 and 0.648, respectively. The users displayed positive results when using Llama3.1:8B in the comparative user evaluation and thought the system reduced the workload significantly. The study concludes that although current models are not accurate enough to fully automate the identification of personal information straight out of the box, integrating local LLMs as an assistive tool shows strong potential and provides value inreal-world GDPR compliance.
Author: Tideman, Oscar
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Umeå Universitet
OpenAlex ID: W7112228961