Source Camera Image Forensics

Abstract

With the advent of low-cost image acquisition devices, storage, and widespread network connectivity, digital images are being increasingly utilized for information capture and dissemination on social media platforms. However, with the widespread use of images there are also increased concerns about the integrity of image information and the misuse of images, for example, for unauthorized information leakage, sharing of illicit photographs, and attacks on privacy. Effective source camera model identification (CMI) techniques can play a crucial role in verifying the trustworthiness and integrity of the digital images and in investigating misuse by locating the source of images. In addition to forensic analysis and image tampering detection, the CMI techniques can also be used for intellectual property protection by identifying the source of copyrighted images, which can help prevent unauthorized use and distribution. This thesis delves into Source Camera Image Forensics (SCIF), a subfield of image source forensics that serves as a blind verification method for digital image authenticity and integrity. SCIF specifically aims to identify individual camera devices and models linked to images. This thesis presents a detailed survey of existing methods for the SCIF. Additionally, it studies and improves the performance of CMI methods, presenting a basic framework for CMI. Within this framework, a dual-branch CNN method is proposed, incorporating improved methodologies for each stage. The first stage involves proposing a patch selection to extract important patches from the input image. In the second stage, high-pass filtering is applied to highlight artifacts related to the camera model, and this filtered image is passed to the second branch of the dual-branch CNN. In the third stage, ResNet is used to extract features from both RGB and high-pass filtered images. The proposed dual-branch CMI method demonstrates significant improvement compared to previous works, compared over multiple datasets. Further, we consider more real-world aspects of images undergoing unknown processing for being shared over social media and making it very challenging to effectively perform CMI on these social media processed images. We present generic and also social media specific CMI models in this thesis. Identifying the Source Social Media Network (SSMN) of the image helps in channeling the image to the respective trained CMI model. So we propose a method, SNRCN2, for identifying the SSMN of digital images. SNRCN2 utilizes high-pass filtered images using steganalysis filters. The experimental results show the superior performance of SNRCN2. Furthermore, motivated by the fact that social media networks apply some Image Processing Operations (IPOs) during image upload, we propose a method, Multi-Scale Residual Deep CNN (MSRD-CNN), for detecting IPOs. Theexperimental results show that MSRD-CNN performs significantly better in classifying images post-processed with various operations. The thesis also acknowledges the growing applications of multispectral images, proposing a novel CMI method tailored for these images. A dual-branch network based on FractalNet rule is introduced, analyzing noise residuals from multispectral channels to classify camera models. Ours is also the first work related to camera identification of multispectral images. Additionally, this thesis introduces a new dataset, IITRPR-CMI, designed to serve as a potential benchmark for evaluating CMI methods. This dataset comprises of a diverse set of images acquired using the contemporary smartphone cameras and features a unique train-test split based on content type and image acquisition methods for better alignment with the real-world scenarios.