Testing human ability to detect ‘deepfake’ images of human faces
The research question explores how well humans can detect deepfakes of human faces from a pool of non-deepfake image and to assess the effectiveness of some interventions intended to improve detection accuracy. Specifically, there were three research questions. Are participants able to differentiate between deepfake and images of real people above chance levels? Do simple interventions improve participants’ deepfake detection accuracy? Does a participants’ self-reported level of confidence in their answer align with their accuracy at detecting deepfakes? The Style Generative Adversarial Network 2 (StyleGAN2) algorithm can produce deepfake images of almost anything, but has been particularly effective at producing deepfake images of human faces. This is the algorithm the researchers used in their study.
The experiment was implemented using a web application and hosted on a secure server. The participants first read about the purpose of the study and the conceptual distinction between images of real and AI-generated faces. Instructions were then given to the participants on how they should complete the experiment. Each participant was shown a sequence of 20 images randomly selected from a pool of 50 deepfake images of human faces and 50 images of real human faces. Participants were asked whether each image was AI-generated or not, to report their confidence, and to describe the reasoning behind each response.
The experiment found that participants were correct at identifying deepfake images 60 percent of the time on average. None of the intervention conditions the researchers were testing as part of their research questions led to a significant improvement in participants’ ability to detect deepfake images. The findings of this study suggest that people are better than random but imperfect at detecting deepfakes and that the simple interventions do not help. A separate study titled “Configuring Fakes: Digitized Bodies, the Politics of Evidence, and Agency” from the Rutgers School of Communication and Information found that deepfakes are disproportionately developed and used to harm women, specifically women of color, LGBT individuals, and those questioning power.
The study investigates the ability of humans to detect deepfake images from similar authentic images, which is likely going to become a major problem for society as the technologies get better. This study provides value that previous studies on this topic did not by having a strong focus on ecological validity and relevance to the wider context of criminal misuse of the technology, and by delving deeper into the participant decision making process when evaluating potential deepfake images. This study incudes topics on emerging technologies impact on society, potential criminal uses of deepfake technology, discussion on cyber education to mitigate cyber related risks (deepfakes), and the psychological factors at play when humans are trying to discern if an image is real or not. These are related to concepts we learned in this class and social science principles.
Main Article Link: https://academic.oup.com/cybersecurity/article/9/1/tyad011/7205694?searchresult=1#415793162
Rutgers Article Link: