TALIU: A Novel Decoder and Augmentation Strategy for Boosting Tampered Document Image Detection

In modern information exchange, document images are vital, often embedding sensitive data. The emergence of advanced image editing tools and generative AI models has elevated the risks associated with document forgery and tampering, representing considerable security concerns. Although many studies have been developed recently, they focus on images saved with compression factors. For lossless compression images like PNG images, current methods based on segmentation show low performance. Thus, this study tackles the specific challenges of identifying tampered text within lossless compression factor images. We introduce a lightweight and efficient model for document image tampering detection, named TALIU, which is based on segmentation networks. Our model employs an encoder-decoder architecture with a novel streamlined iterative upsampling decoder specifically designed for the context of document tampering. Moreover, we present an innovative data augmentation strategy, tampered region augmentation, intended to enhance sample diversity during the training phase. Extensive experiments utilizing the DocTamper dataset, the largest of its kind, show that our TALIU model surpasses current state-of-the-art methods in both detection accuracy and computational efficiency for the problem of tampered text in lossless compression factor images.