Please use this identifier to cite or link to this item:
Title: Passive approaches for digital image forgery detection
Authors: Pravin Kakar
Keywords: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
Issue Date: 2012
Source: Pravin, K. (2012). Passive approaches for digital image forgery detection. Doctoral thesis, Nanyang Technological University, Singapore.
Abstract: Fake images have become widespread in society today. One can find forged images used to sensationalize news, spread political propaganda and rumors, introduce psychological bias, etc. in all forms of media. Claims of image tampering are common in scandals and controversies. As the credibility of images suffers, it is necessary to devise techniques in order to verify their authenticity. Traditionally, contextual knowledge and/or active approaches such as watermarking or signatures have been employed. However, contextual knowledge is not always avail- able or applicable, and active approaches have shortcomings such as the requirement of insertion at the point of creation of the image or later by an authorized agent. In the face of such difficulties in authenticity verification, passive approaches have gained importance as they only use the characteristics of the image in question in order to establish its genuineness. One of the novel techniques that we have developed for the passive detection of image forgeries utilizes discrepancies in motion blur in order to detect spliced images. We estimate the motion blur in a blockwise manner, using either spectral matting or image gradients, over the entire image, and then determine regions where the blur is inconsistent with the rest of the image. In images with visually similar blur, this can be an indication of the possible presence of splicing. The technique that we have developed is the first to use motion blur to find inconsistencies in images, and advances the state of the art beyond defocus blur-based techniques. Another technique that we have proposed uses the MPEG-7 visual signature tools in order to detect copied regions in the same image. These tools have been shown to be robust to various common image processing operations in the context of content-based image retrieval (CBIR). We have adapted these tools to work with image forgeries, where the set of problems faced is different from CBIR. We have obtained very good results with extremely low false positives for tampered images subjected to many common image processing operations. As compared to the state of the art in image forensics, our method is able to handle more postprocessing operations and performs at a much lower false positive rate while still maintaining a high true positive rate. Our third proposed technique uses the dependence of the position of the sun in the sky on location, date and time in order to verify image temporal metadata. For images containing geolocation information in their metadata, we utilize ground shadows to estimate the azimuthal direction of the sun, and compare it against a calculated theoretical value in order to authenticate the time of capture of the image. Through this technique, we have enabled the authentication of the time of capture of an image, which was previously restricted to a much coarser day or month of capture. We have proposed a novel two-stage shadow detection technique in order to reliably detect ground shadows. We have also incorporated the use of vertical surfaces and the sky in order to improve the reliability of the estimated sun direction. Lastly, we have also discussed other applications of our technique such as verifying the date of capture of an image and determining the camera direction at the time of capture. Thus, in this dissertation, we have proposed methods to detect three different types of image forgeries in a passive manner - splicing, copy-paste, and metadata tampering.
DOI: 10.32657/10356/50595
Fulltext Permission: open
Fulltext Availability: With Fulltext
Appears in Collections:SCSE Theses

Files in This Item:
File Description SizeFormat 
low_res.pdfThesis5.65 MBAdobe PDFThumbnail

Page view(s) 5

Updated on Jan 15, 2022

Download(s) 1

Updated on Jan 15, 2022

Google ScholarTM




Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.