Can people identify original and manipulated photos of real-world scenes?

Abstract

Advances in digital technology mean that the creation of visually compelling photographic fakes is growing at an incredible speed. The prevalence of manipulated photos in our everyday lives invites an important, yet largely unanswered, question: Can people detect photo forgeries? Previous research using simple computer-generated stimuli suggests people are poor at detecting geometrical inconsistencies within a scene. We do not know, however, whether such limitations also apply to real-world scenes that contain common properties that the human visual system is attuned to processing. In two experiments we asked people to detect and locate manipulations within images of real-world scenes. Subjects demonstrated a limited ability to detect original and manipulated images. Furthermore, across both experiments, even when subjects correctly detected manipulated images, they were often unable to locate the manipulation. People's ability to detect manipulated images was positively correlated with the extent of disruption to the underlying structure of the pixels in the photo. We also explored whether manipulation type and individual differences were associated with people's ability to identify manipulations. Taken together, our findings show, for the first time, that people have poor ability to identify whether a real-world image is original or has been manipulated. The results have implications for professionals working with digital images in legal, media, and other domains.

Keywords: Digital image forensics; Photo manipulation; Psychology and law; Real-world scenes; Visual processing.

Fig. 1

Samples of manipulated photos. a Original photo; b airbrushing—removal of sweat on the nose, cheeks, and chin, and removal of wrinkles around the eyes; c addition or subtraction—two links between the columns of the tower of the suspension bridge removed; d geometrical inconsistency—top of the bridge is sheered at an angle inconsistent with the rest of the bridge; e shadow inconsistency—face is flipped around the vertical axis so that the light is on the wrong side of the face compared with lighting in the rest of the scene; f super-additive—combination of all previously described manipulations. Original photo credit: Vin Cox, CC BY-SA 3.0 license. Photos b–f are derivatives of the original and licensed under CC BY-SA 4.0

Fig. 2

Example of a photo with the location grid overlaid. Photo credit: Vin Cox, CC BY-SA 3.0 license

Fig. 3

Mean proportion of correct “detect” and “locate” decisions by type of photo manipulation. The dotted line represents chance performance for detection. The grey dotted lines on the locate bars represent chance performance by manipulation type in the location task. Error bars represent 95% CIs

Fig. 4

Mean proportion of correct “locate” decisions when subjects correctly detected that the photo was manipulated (i.e., correctly said “Yes” on the detection task). The grey dotted lines on the bars represent chance performance for each manipulation type. Error bars represent 95% CIs

Fig. 5

Mean proportion of correctly detected (a) and located (b) image manipulations by extent of pixel distortion as measured by Delta-E. The graphs show individual data points for each of the 30 manipulated images

Fig. 6

Mean proportion of correctly detected (a) and located (b) image manipulations by extent of pixel distortion as measured by Delta-E. The graphs show the mean values for each of the five categories of manipulation type

Fig. 7

Mean proportion of manipulated photos accurately detected and accurately located (DL), accurately detected, inaccurately located (DnL), inaccurately detected, accurately located (nDL), and inaccurately detected, inaccurately located (nDnL) by manipulation type. The dotted horizontal lines on the bars represent chance performance for each manipulation type from the results of the Monte Carlo simulation. The full horizontal lines on the bars represent chance performance for each manipulation type based on subjects’ responses on the original image trials. Error bars represent 95% CIs

Fig. 8

Mean proportion of correctly detected (a) and located (b) image manipulations by extent of pixel distortion as measured by Delta-E. The graphs show individual data points for each of the 30 manipulated images

Fig. 9

Mean proportion of correctly detected (a) and located (b) image manipulations by extent of pixel distortion as measured by Delta-E. The graphs show the mean values for each of the five categories of manipulation type