SNRCN2: Steganalysis noise residuals based CNN for source social network identification of digital images

Abstract

Identifying the processing history of digital images related to source acquisition device, image manipulations, and Source Social Media Network (SSMN) identification is important for a forensic analyst to verify image source, trustworthiness, and integrity. Nowadays, social media networks have become a major medium of image sharing and identifying the SSMN of digital images has allured attention in the image forensic scientific community. With the precedence of deep Convolutional Neural Networks (CNNs) in the domain of image forensics, we propose a novel Steganalysis Noise Residuals based CNN (SNRCN2) for digital images SSMN identification. Inspired by the fact that image content can heavily obscures the artifacts of post-processing operations, the proposed CNN utilizes the features from steganalysis-based noise residuals to highlight the artifacts introduced by social media networks. Therefore, the given input image is high-pass filtered using well-known 30 Spatial Rich Model (SRM) filters to obtain noise residuals. Afterward, these noise residuals are passed to an efficient CNN for the extraction of high-level features related to social media networks. Lastly, a fully-connected layer is used for classification purposes. The experimental results show that the proposed model outperforms the existing techniques by providing an image-level accuracy of 99.53% and 100% on the VISION and Forchheim datasets, respectively. To further evaluate the robustness of the proposed model, we create a combined dataset that includes the images of common classes from both datasets. The results of the combined dataset further confirm the efficacy of the proposed model.