Image manipulations analysis and detection methods for reflection-based attacks / Nor Bakiah Abd. Warif

• Main Author: Nor Bakiah, Abd. Warif
• Format: Thesis
• Published: 2018
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://studentsrepo.um.edu.my/8681/1/Nor_Bakiah.pdf; http://studentsrepo.um.edu.my/8681/6/bakiah.pdf; http://studentsrepo.um.edu.my/8681/
• Institution: Universiti Malaya

The extensive use of digital images at this age has led to the content manipulations that misrepresent the information with malicious intent. This issue demands the requirement of digital image investigation to verify the sources and validate trustworthiness. One of the image manipulation techniques is called copy-move forgery (CMF). It is a process of duplicating one or more regions in an image before being pasted to another location within the same image. The CMF is mainly comprised of translation attacks and commonly combined with other attacks, such as scaling, rotation, compression and Gaussian noise addition. In this thesis, the research is divided into two stages. The first stage looks into the performance analysis of the existing CMF detection methods, while the second stage focuses on proposing detection methods for reflection-based attacks in CMF. At present, CMF detection performance is evaluated, either through image-level evaluation, pixel-level evaluation, or both. Since there is no evaluation standard, the analysis also studies the effects of these evaluations towards the result interpretation. The study shows that both image and pixel-level evaluations are dependent, therefore, must be incorporated together to ensure fair evaluation. These evaluations are then applied to study the effects of reflection-based attacks in the second stage of research. Methods called SIFT-Symmetry and CMF-iteMS are proposed to alleviate the reflection-based problems in CMF. The SIFT-Symmetry incorporates symmetry matching in a keypoint-based CMF detection while the CMF-iteMS uses a block-based approach that includes iterative means of region size. To evaluate the performance of the two proposed methods, they are compared with state-of-the-arts methods, based on keypoint, block, and a combination of both approaches. The evaluations involve CombineTranslation, CPHALL, NB-Casia, and NBr-Casia datasets which include translation, scale, rotation, and reflection attacks. The results are measured using multiple F-score values which are for image, pixel, and both, image and pixel. The image score shows the ability of the detection methods in distinguishing the original image and CMF image, while the pixel score defines the reliability of determining the exact location of the CMF detection. Both scores are multiplied to get the overall percentages of the detection. The CMF-iteMS surpassed the minimum value of 96% for image score and 88% for both pixel score and percentages of detection for simple translation attacks, while having maintained the highest percentages for rotation, simple reflection, and a mix of attacks with the minimum value of 87%, 76%, and 62%, respectively. In terms of reflection-based CMF, the CMF-iteMS achieved the highest percentages in all reflection cases even if the reflection is combined with scale attacks. Alternatively, the SIFT-Symmetry obtained the highest image score with a value of 94% for simple reflection and 75% for reflection with scale attacks. Moreover, the results also proved that the combination of the existing CMF detection methods with the iterative means of region size increases the performance of the block-based approach. The combination with other approaches, on the other hand, is able to reduce the spurious matching even though the percentages of both image and pixel-levels are dropped.