CONAN: Markov Image-Based Preprocessing for Robust Screenshot Detection Using Deep Learning

The development of information technologies has facilitated the generation and sharing of creative works, contributing to the advancement of intellectual property. However, at the same time, with the widespread distribution of intellectual property, malicious users are indiscriminately sharing content through methods such as taking unauthorized screenshots. To address these copyright infringement issues, several studies have previously been conducted on detecting the metadata of screenshot images or analyzing surface-level features, such as structural elements like user interface layouts or edge patterns. However, these metadata and surface-level features can be easily manipulated in screenshot images to evade detection. It is necessary to develop a more robust detection method to distinguish screenshot images from original (i.e., non-screenshot) images. In this paper, we propose a novel classification method, named CONAN (Classification Of origiNal And screeNshot images), to detect screenshot images from original images using byte-level features. CONAN converts images into Markov images as a preprocessing step, and uses them to train a deep learning model. As Markov images can capture the byte-level features of screenshot images, CONAN can accurately and robustly distinguish the screenshot images from the original images. Our empirical evaluation shows that the accuracy in classifying original and screenshot images reached 99.8%, significantly outperforming the existing techniques. We also validated that CONAN’s preprocessing technique can be robustly applied to various deep learning models and under diverse conditions.