Unmask Tampering: Efficient Document Tampering Localization under Recapturing Attacks with Real Distortion Knowledge

Abstract

Document tampering localization (DTL) aims to detect tampering traces and ensure the integrity of document images. However, recapturing attacks (i.e., printing and scanning the altered document images) can effectively conceal tampering traces due to distortions such as halftoning, blurring, and noise. Compounding this challenge, the collection of real recaptured document samples is both time-consuming and resource-intensive. It is important to investigate an efficient method that adapt the existing DTL models to unmask the threat from recapturing attack. In this work, we tackle these challenges by first proposing a Real Halftone-based Document Synthesis (RHSyn) method to generate realistic recaptured document images. RHSyn exploits reference halftone patterns with a novel table look-up operation, which incorporates real-world distortions from printers and scanners to produce high-fidelity synthetic data. To improve DTL performance, we introduce a Masked Parameter-Efficient Fine-Tuning (M-PEFT) technique to facilitate extracting distinctive forensic features from text and background regions under recapturing attacks. In the experiment, we gather two extensive testing datasets comprising over 6,600 real recaptured document images from 9 printers and 7 scanners. Experimental results under recapturing attacks demonstrate that the performances of the existing DTL models are significantly improved with RHSyn-generated data via M-PEFT. Specifically, our approach achieves an average F1-Score of 0.611 across three test datasets, increased by 0.496 compared to models without fine-tuning, demonstrating its capacity to effectively counter the threat of recapturing attacks.