Seite - 169 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure3: CreationofNewKeypoints. 3x3 Patch 5x5Patch 7x7Patch 9x9Patch 43.01% 39.76% 34.43% 32.21% Table2: NewlyGeneratedKeypoints inCA. Whennewkeypoints arebeinggenerated, it isnot their number that is most important. The aim of re- moving keypoints is compromised, if thenewlygen- erated ones are similar to the removed ones in terms of their SIFT descriptors. In this case, attacks might stillberecoveredbytheforensicanalysteventhough keypoints have been removed. In Fig. 4 we plot the distance (squared Euclidean distance (SED)) of the SIFT descriptors describing removed and newly cre- ated ones. In particular, we compute SED between removedkeypointsand theirclosestnewlygenerated keypoints in termsof theirdescriptors. Toavoidbias, we divide the result by the number of removed key- points, as we display results in terms of increasing percentage of removedkeypoints. For the patch-based techniques, an increase of the patch size leads to higher SED, which is expected and desired. When increasing the percentage of re- moved keypoints, there is a tendency for increasing SED,except forLSandCAwithsmallerpatchsizes. The largest SED values (which is the aim when re- moving keypoints) are seen for techniques involving GS (not shown) when a large share of all keypoints is being removed. CA clearly exhibits the lowest values, which means that the advantage of this ap- proach in removing all keypoints is endangered by the creation of new keypoints which are close to the 0 10000 20000 30000 40000 50000 60000 5 10 15 20 25 30 35 40 45 50 55 60 65 70 Keypoints Removed ( in Percent) Distance between Disciptors (Delected & Newly Created) 3x3 Patch 5x5 Patch 7x7 Patch 9x9 patch (a)LS 0 5000 10000 15000 20000 25000 30000 35000 40000 5 10 15 20 25 30 35 40 45 50 55 60 65 70 Keypoints Removed ( In Percent) Distance Between Discriptors (Deleted & Newly Created) 3x3 Patch 5x5 Patch 7x7 Patch 9x9 patch (b)CA Figure4: Distance tonewlycreatedkeypoints. removedones in termsof theirSIFTdescriptors. After having analysed four different SIFT key- point removal techniques with respect to different properties,wetested thesemethods inanactualcopy moveforgeryscenario. Thefollowingdefinitionsare employed: • TruePositive(TP): A true positive test result for a forged image is one that detects at least τ matchingkeypointpairs. • FalseNegative(FN): Afalsenegative test re- sultforaforgedimageisonethatdetectsatmost τ−1matchingkeypointpairs. • TrueNegative(TN): A true negative test re- sult for an image fromtheBUILDING PHOTO categoryisonethatdetectsatmostτ−1match- ingkeypointpairs. 169