Dual-Color Space Hierarchical Classification Network for Image-Sharing Chain Detection

Numerous controversial images spread across online social networks (OSNs) and identifying the sources of these images is crucial. Different OSNs manipulate uploaded images in various ways to reduce transmission and storage costs, leaving identifiable traces. Sharing chain detection aims to reconstruct the sharing history of images on OSNs based on these traces. This article proposes a new sharing chain detection method called the dual-color space hierarchical classification network. We first use the RGB and YCbCr color spaces to maximize perceiving traces within the image pixels, and we also employ an adaptive patch selection (APS) module to choose the most suitable patch for the network input. Furthermore, we innovatively use a hierarchical classification approach to address the sharing chain problem, treating it as a temporal process rather than a simple category. We use multiple interconnected classifiers to detect OSNs sequentially, with classifiers accepting features at different stages, from shallow to deep. The experimental results show that the proposed method can effectively detect the sharing chain of up to three different OSNs. Compared to the state-of-the-art method, the accuracy of detecting the last two OSNs has improved by 2.20% on R-SMUD and 2.09% on V-SMUD.