SIFT-Symmetry: A robust detection method for copy-move forgery with reflection attack

Highlights

• CMF detection with rotation and scale as well as reflection-based attack is studied.

• The purpose is to achieve high robustness against CMF with reflection and any combination attacks.

• SIFT-Symmetry is a combination of SIFT-based CMF detection with symmetry-based matching.

• SIFT-Symmetry is evaluated on two new datasets and compared with three methods.

• SIFT-Symmetry’s F-score is superior than other existing methods.

Abstract

Copy-move forgery (CMF) is a popular image manipulation technique that is simple and effective in creating forged illustrations. The bulk of CMF detection methods concentrate on common geometrical transformation attacks (e.g., rotation and scale) and post-processing attacks (e.g., Joint Photographic Experts Group (JPEG) compression and Gaussian noise addition). However, geometrical transformation that involves reflection attacks has not yet been highlighted in the literature. As the threats of reflection attack are inevitable, there is an urgent need to study CMF detection methods that are robust against this type of attack. In this study, we investigated common geometrical transformation attacks and reflection-based attacks. Also, we suggested a robust CMF detection method, called SIFT-Symmetry, that innovatively combines the Scale Invariant Feature Transform (SIFT)-based CMF detection method with symmetry-based matching. We evaluated the SIFT-Symmetry with three established methods that are based on SIFT, multi-scale analysis, and patch matching using two new datasets that cover simple transformation and reflection-based attacks. The results show that the F-score of the SIFT-Symmetry method surpassed the average 80% value for all geometrical transformation cases, including simple transformation and reflection-based attacks, except for the reflection with rotation case which had an average F-score of 65.3%. The results therefore show that the SIFT-Symmetry method gives better performance compared to the other existing methods.

Introduction

Copy-move forgery (CMF) is an image manipulation technique that is commonly used for information exploitation due to its simplicity and highly visual effects. Also known as region duplication or cloning, CMF copies a region of an image and moves that region to another region within the same image. The technique may be used to give a false impression to favor an individual’s personal agenda. This includes turning a small crowd into a larger crowd to show the impression of great support or hiding objects in an image to conceal unwanted information. Owing to the damaging effects that can be caused by the misleading information, several methods have been proposed to detect CMF in an image [1], [2], [3], [4], [5], [6], [7]. The proposed CMF detection methods are normally evaluated by looking at their sensitivity towards geometrical transformation attacks and post-processing attacks. For a geometrical transformation or intermediate attack [8], resilience against spatial manipulation activities such as rotation, scale, and reflection are evaluated. Conversely, researchers have investigated the robustness against the blending effect that reduces visual manipulation footprints through post-processing attacks. Such attacks include compression, noise, and blurring effects are occur after the geometrical transformation manipulation.

The proliferation of image editing tools, i.e., Adobe Photoshop, Paint, and Microsoft Office, that can perform CMF has motivated us to analyze common geometrical transformation attacks such as rotation and scale, as well as the mixture between the two in this research. Special attention is given to reflection-based transformation attack that has not been highlighted by prior researchers. As reflection could change the feature organization of an image, further investigation on such attacks should be explored. The changes in the feature organization may also affect the performance of typical CMF detection methods [5].

We start our investigation by collecting and creating suitable data that contain images for simple transformation attacks and reflection-based attacks. The original data are taken from CASIA v2.0, and the simple transformation attacks are divided into five groups, namely, simple translation, simple scale, simple rotation, simple reflection, and the mixture of simple translation, scale, and rotation. In contrast, reflection-based data contain simple reflection, reflection with scale, reflection with rotation, and the combination of reflection with scale and rotation manipulation. We then establish our method called SIFT-Symmetry that emphasizes reflection-based attacks. Our method combines the Scale Invariant Feature Transform (SIFT)-based CMF method with the symmetry matching technique. We then compare our performance by F-scores with three state-of-the-art methods: Amerini et al.’s method [5], Cozzolino et al.’s method [6], and Silva et al.’s method [7].

This article is divided into five parts. Section 2 presents related works on CMF detection per the attacks involved. Section 3 explains the proposed method for CMF detection. Section 4 discusses the experimental results and evaluation. Section 5 presents the conclusion and future studies.