Anti-Screenshot Watermarking Algorithm for Archival Image Based on Deep Learning Model

Abstract

Over recent years, there are an increasing number of incidents in which archival images have been ripped. Leak tracking is one of the key problems for anti-screenshot digital watermarking of archival images. Most of the existing algorithms suffer from low detection rate of watermark, because the archival images have a single texture. In this paper, we propose an anti-screenshot watermarking algorithm for archival images based on Deep Learning Model (DLM). At present, screenshot image watermarking algorithms based on DLM can resist screenshot attacks. However, if these algorithms are applied on archival images, the bit error rate (BER) of the image watermark will increase dramatically. Archival images are ubiquitous, so in order to improve the robustness of archival image anti-screenshot, we propose a screenshot DLM “ScreenNet”. It aims to enhance the background and enrich the texture with style transfer. Firstly, a preprocessing process based on style transfer is added before the insertion of an archival image into the encoder to reduce the influence of the screenshot process of the cover image. Secondly, the ripped images are usually moiréd, so we generate a database of ripped archival images with moiréd by means of moiréd networks. Finally, the watermark information is encoded/decoded through the improved ScreenNet model using the ripped archive database as the noise layer. The experiments prove that the proposed algorithm is able to resist anti-screenshot attacks and achieves the ability to detect watermark information to leak the trace of ripped images.

A Brief, In-Depth Survey of Deep Learning-Based Image Watermarking

“…One intricate area involves extracting watermarks from images that have undergone resampling via a camera, which introduces multifarious noise types including JPEG artifacts, variations in lighting, and optical distortions. In response to this, Fang et al [33] and Gu et al [34] advanced approaches that incorporate a screen-shooting noise layer simulation. This adaptation enables the simulation of camera resampling noises like geometric distortions, optical bends, and RGB ripples within the training of deep learning-based image watermarking models, fostering a more robust system capable of counteracting these specific noise introductions.…”

On the Role and Form of Personal Information Disclosure in Cyberbullying Incidents

“…One potential direction is watermarking [49] personal information on OSN platforms in order to track and possibly block the relaying of personal information captured. Such a mechanism should mark personal information directly (e.g., photos) or indirectly (screenshots of conversations) by including the watermark in the UI of the app [32,50,70], thus enabling the tracking of the information and determining whether it has been disclosed with or without the consent of a data subject (i.e., the potential victim). If implemented successfully, this approach could make disclosers accountable and could reduce the risk of using personal information for CB.…”