Page - 166 - in Joint Austrian Computer Vision and Robotics Workshop 2020

EvaluatingCounterMeasuresagainstSIFTKeypointForensics MuhammadSalman,AndreasUhl DepartmentofComputerSciences,UniversityofSalzburg uhl@cs.sbg.ac.at Abstract. Forensic analysis is used to detect image forgeries e.g. the copy move forgery and the object removal forgery. Counter forensic techniques (meth- ods to fool the forensic analyst by concealing traces of manipulation) have become popular in the game of cat and mouse between the analyst and the at- tacker. Methods tocounter forensic techniquesbased on SIFT keypoints are being analysed in this paper (aka anti-forensic techniques), with particular em- phasis on keypoint removal in the context of copy move forgery detection. Local smoothing is sug- gested in this paper and turns out to be a highly at- tractive alternative to techniques investigated in lit- erature so far. 1. Introduction In the past, images were considered as an authen- ticsourceofinformation–withincreasingpopularity and the availability of low-cost image editing soft- ware such as Adobe photoshop, corel paint shop and GIMP the truthfulness of an image can no longer be taken for granted. Among other forgery types, copy move forgery and object removal forgery are the most prominent ones. In a copy move forgery, a part of the image itself is copied and pasted into another part of the same image to conceal an impor- tant object or information, or to conceal that an ob- ject has been removed from the image in an object removal forgery. Inmostcasesof imageforgery, it is extremelydifficult todistinguishbetweenanoriginal imageandtheforgedone. Therefore, it is required to develop methods/techniques to assess the authentic- ityofan image–Digital ImageForensics (DIF[19]) has served this purpose to a large extent. Whenever an image is forged, there are some traces which are leftbehind in theforged image. These tracesareuse- ful for the forensic researcher todetect a forgery. Awide rangeofDIF forgerydetection techniques have been established in the recent years [4, 6, 21]. Besides recent deep learning based schemes, tech- niques relying on Scale Invariance Feature Trans- form (SIFT) keypoints have been shown to be effec- tive. In particular, SIFT keypoints [12] have been proposed to reveal copy move forgeries [6] and im- agecloning[17],aswellas todetectcopyrightedma- terial usingCBIRtechniques [9]. Attackers are making it difficult to apply these techniques by developing counter forensic tech- niques, i.e. by minimising those traces left behind in forged images. In the context of SIFT keypoint forensics, this is done by manipulating SIFT key- points, e.g. removing existing ones or injecting fake keypoints tofool theforensic techniques. Thispaper is acontribution to suchcounter forensicapproaches againstSIFT-keypoint forensic techniques. Inpartic- ular,wefocusonSIFTkeypoint removal techniques. Section 2 reviews corresponding techniques as pro- posed in literature and suggest a new approach. Sec- tion 3 is devoted to an extensive empirical evalua- tion, looking at the tradeoff among image quality, keypoint removal effectiveness as well as the gener- ationofnewkeypoints. In theconclusionwediscuss results obtained and give an outlook to further work in thisdirection. 2.SIFTKeypointRemovalTechniques The simplest approach, global smoothing (GS), reduces the potential keypoints at the level of dif- ference of Gaussian (DoG) by Gaussian smoothing (whichflattens thepixelvaluesofan image), e.g. [1] applies a Gaussian filter with σ = 0.7 and window size 3× 3 as a good compromise between amount of deleted keypoints and overall visual quality of an image. A more sophisticated approach is to first ap- ply GS (the original paper [9] suggests to employ σ = 1.3), detect remaining keypoints, and apply lo- cal smoothing (LS) in patches around detected key- 166