Source Camera Identification for Compressed JPEGs through PRNU and Machine Learning

Abstract

Online Social networks (OSN) such as Facebook, Twitter, and WhatsApp, etc. are excessively used to socialize or gain insights into people’s behavior and their activities. Everyday millions of images and videos are captured and uploaded, which is then misused by a felon to fulfill their evil-deeds. Today’s digital cameras store image metadata along with the image itself on the media card, DVDs, and RD-ROMs. The image metadata includes the data about the image such as the date and time the image was captured, the focal length, shutter stops, make, and model of the camera. If a forensics examination is conducted only on the computer then metadata can be easily used to identify the images and to link back to a specific camera. However, the metadata can be easily editable through different software or can be lost if the image file format changes, for example, software processing is applied, downloaded, or transfer through online social networks. It is difficult to standardize digital images for the courtroom due to rapid changes in the growth of computer-related technologies. The Digital Forensic Research community keeps suggesting innovative approaches to study digital videos and images and also to identify the probable source. This can be done by analyzing the camera sensor-specific artifacts left behind in an image. However, the high degree of compression destroys the camera sensor-specific artifacts by uploading the image to online networking sites. JPEG compression is one of the standard processes in most consumer-level cameras. It is a standard compression algorithm, however, the size and the quality tradeoff are at the manufacturer’s and the users’ discretion. Our work is to attempt to investigate the source camera identification problem, we present a model on an unknown model detection problem to find the source of the image through source camera identification that works on compressed images too. We propose a PRNU and machine learning-based method to identify the source with an accuracy of 98%. We perform our experiments using images from the Dresden dataset. The proposed approach outperforms the current state-of-the-art in terms of accuracy in source camera identification on JPEG compressed images.