Comparative Study between Leading Transfer Learning Architectures for Source Camera Identification

The all-embracing use of digital images has revamped the quality of life and security to a great extent. Right from finding an item on online shopping websites through a clicked picture, to CCTV cameras being used for road traffic control, the users have learnt to appreciate the existence of technology being as advanced. However, one cannot overlook the gravity of this technology being misused. Although, the digitization has incorporated advanced concepts like Computer Vision and Deep Learning for security-check and crowd control, this has encouraged the advancement of courtroom discussions. Framing people for wrongdoings they are not involved with, on the basis of a fake “digital proof,” is one of the newly faced muddles. False allegations on a person, on the basis of a picture or a video, can potentially put a question on the existence of a person. The need to find the legitimacy of a produced image is therefore, of utmost importance. There have been various studies over the years, wherein a lot of methods were proposed to develop a system that identifies the camera model. Through this paper, we aim to produce a comparative study between four leading architectures, DenseNet, Inception V3, MobileNetV2 and Exception(XCeption), and suggest a the most competent architecture for commercialization of this system.