Detection and Anonymisation of Sensitive Data in Financial Invoices

Sammanfattning

The increasing use of artificial intelligence (AI) for processing large volumes of business documents has created a growing need to handle sensitive information securely and in compliance with privacy regulations. Financial invoices often contain personally identifiable information (PII), such as personal identification numbers, names, addresses, email addresses, and phone numbers. Due to their diverse layouts and multimodal nature, invoices present a significant challenge for automated detection and anonymisation of sensitive data. This project investigates how accurately sensitive information in such invoices can be detected and anonymised using methods that require minimal setup. Addressing this problem is important, as improper handling of PII can lead to privacy violations, legal consequences under regulations such as the General Data Protection Regulation (GDPR), and limitations in the use of real-world data for AI development. To address this challenge, a system was developed using a design science research (DSR) approach to explore and evaluate multiple anonymisation techniques. The evaluated methods included rule-based approaches using regular expressions, a large language model (GPT-5.1), a layout-aware structured document understanding model (LiLT), and combinations of these methods. The system was tested on invoice datasets in both PDF and image formats, and performance was evaluated using precision, recall, and F1-score metrics. The results indicate that a combination of LiLT and rule-based methods was the most effective approach for anonymisation. The proposed solution achieved a precision of 0.825, a recall of 0.886, and an F1-score of 0.839. These findings suggest that combining complementary methods provide promising anonymisation performance for invoices, even when document layouts vary.