Discovering Transferable Forensic Features for CNN-Generated Images Detection

AbstractVisual counterfeits (We refer to CNN-generated images as throughout this paper.) are increasingly causing an existential conundrum in mainstream media with rapid evolution neural image synthesis methods. Though detection of such has been a taxing problem the forensics community, recent class forensic detectors – universal able surprisingly spot counterfeit regardless generator architectures, loss functions, training datasets, and resolutions [61]. This intriguing property suggests possible existence transferable features (T-FF) detectors. In work, we conduct first analytical study discover understand T-FF Our contributions 2-fold: 1) We propose novel feature relevance statistic (FF-RS) quantify and, 2) qualitative quantitative investigations uncover unexpected finding: color is critical Code models available at https://keshik6.github.io/transferable-forensic-features/.