Anonymization of Documents for Law Enforcement with Machine Learning

TL;DR

The paper tackles the challenge of protecting personal data in law-enforcement contexts by automating the anonymization of scanned documents. It couples region detectors for faces, text, barcodes, MRZs, and signatures with a reference-based redaction transfer pipeline guided by instance retrieval from a single anonymized exemplar, using a self-supervised DinoV2 model to locate the appropriate reference. An affine alignment, driven by A-KAZE keypoints and RANSAC, maps reference redactions to the target document, with per-class rules (notably for text width) to accommodate layout differences. On a proprietary multi-document dataset, the approach significantly outperforms purely automatic redaction and naive reference-copy baselines, especially for text regions, while acknowledging limitations in signature detection and OCR-dependent cases. The framework reduces manual effort and enables scalable, privacy-preserving processing across governmental and private sectors, though it emphasizes the need for human verification and continuous dataset/model improvements.

Abstract

The steadily increasing utilization of data-driven methods and approaches in areas that handle sensitive personal information such as in law enforcement mandates an ever increasing effort in these institutions to comply with data protection guidelines. In this work, we present a system for automatically anonymizing images of scanned documents, reducing manual effort while ensuring data protection compliance. Our method considers the viability of further forensic processing after anonymization by minimizing automatically redacted areas by combining automatic detection of sensitive regions with knowledge from a manually anonymized reference document. Using a self-supervised image model for instance retrieval of the reference document, our approach requires only one anonymized example to efficiently redact all documents of the same type, significantly reducing processing time. We show that our approach outperforms both a purely automatic redaction system and also a naive copy-paste scheme of the reference anonymization to other documents on a hand-crafted dataset of ground truth redactions.

Figures (2)

• Figure 1: The overview of our proposed framework: The input document is used to retrieve already correctly anonymized documents. Various object detection algorithms predict where text, images of faces, the machine-readable zone or barcodes are located on the given document. Afterwards, the predicted bounding boxes are filtered with the reference document and adjusted, e.g. the box width and height. Finally, object types that were not recognized are transferred from the reference document to the input document by using affine transformations. The input document in this figure is a specimen document provided by the German government and does not refer to a real person.
• Figure 2: Four images of different document types and the anonymization of the different methods. On the left, we show the ground truth, followed by the two baseline methods and our proposed method. The color of the redacted bounding box indicates the different elements recognized. Yellow boxes are faces, green indicates text, blue for signatures, black for the mrz and red for barcodes. The turquoise color shows the difference between the predicted bounding box and the ground truth. From top to bottom, we show the front of an Italian driver's license, a German bank transfer form, the front of a Bulgarian ID card and the back of a Polish driver's license.