Where Do Deep Fakes Look? Synthetic Face Detection via Gaze Tracking.

摘要: Following the recent initiatives for democratization of AI, deep fake generators have become increasingly popular and accessible, causing dystopian scenarios towards social erosion trust. A particular domain, such as biological signals, attracted attention detection methods that are capable exploiting authenticity signatures in real videos not yet faked by generative approaches. In this paper, we first propose several prominent eye gaze features fakes exhibit differently. Second, compile those into analyze compare videos, formulating geometric, visual, metric, temporal, spectral variations. Third, generalize formulation to problem a neural network, classify any video wild or real. We evaluate our approach on datasets, achieving 89.79\% accuracy FaceForensics++, 80.0\% Deep Fakes (in wild), 88.35\% CelebDF datasets. conduct ablation studies involving different features, architectures, sequence durations, post-processing artifacts. Our analysis concludes with 6.29\% improved over complex network architectures without proposed signatures.