MDM-CPS: A few-shot sample approach for source camera identification

Abstract

The purpose of source camera identification (SCI) is to identify the source device of target images, so as to ensure the source reliability of digital images. However, most state-of-the-art results require sufficient training samples which are hard to obtain in practice. In this work, we propose an approach based on multi-distance measures and coordinate pseudo-label selection (MDM-CPS) approach to solve the problem of few-shot sample databases. Based on semi-supervised learning, this approach iteratively expands and updates the labeled database. Our approach drastically reduces the interference of noisy pseudo-labels in training and ensures highly-confident prediction of the pseudo-label samples. Through comprehensive experiments, our approach has achieved the best performance in few-shot sample scenarios of the common benchmark databases (i.e., Dresden database and VISION database) in the field of source camera identification.