File(s) under permanent embargo
Provenance analysis for instagram photos
conference contributionposted on 2019-01-01, 00:00 authored by Y Quan, Xufeng Lin, Chang-Tsun LiChang-Tsun Li
© Springer Nature Singapore Pte Ltd. 2019. As a feasible device fingerprint, sensor pattern noise (SPN) has been proven to be effective in the provenance analysis of digital images. However, with the rise of social media, millions of images are being uploaded to and shared through social media sites every day. An image downloaded from social networks may have gone through a series of unknown image manipulations. Consequently, the trustworthiness of SPN has been challenged in the provenance analysis of the images downloaded from social media platforms. In this paper, we intend to investigate the effects of the pre-defined Instagram images filters on the SPN-based image provenance analysis. We identify two groups of filters that affect the SPN in quite different ways, with Group I consisting of the filters that severely attenuate the SPN and Group II consisting of the filters that well preserve the SPN in the images. We further propose a CNN-based classifier to perform filter-oriented image categorization, aiming to exclude the images manipulated by the filters in Group I and thus improve the reliability of the SPN-based provenance analysis. The results on about 20, 000 images and 18 filters are very promising, with an accuracy higher than 96% in differentiating the filters in Group I and Group II.