Multi-Scale Enhanced Dual-Stream Network for Facial Attribute Editing Localization

The advancement of Facial Attribute Editing (FAE) technology allows individuals to effortlessly alter facial attributes in images without discernible visual artifacts. Given the pivotal role facial features play in identity recognition, the misuse of these manipulated images raises significant security concerns, particularly around identity forgery. While existing image forensics algorithms primarily concentrate on traditional tampering methods like splicing and copy-move and are often tailored to detect tampering in natural landscape images, they fall short in pinpointing FAE manipulations effectively. In this paper, we introduce two FAE datasets and propose the Multi-Scale Enhanced Dual-Stream Network (MSDS-Net) specifically for FAE Localization. Our analysis reveals that FAE artifacts are present in both the spatial and DCT frequency domains. Uniquely, in contrast to traditional tampering methods where modifications are localized, facial attribute alterations often span the entire image. The transitions between edited and unedited regions appear seamless, devoid of any conspicuous local tampering signs. Thus, our proposed method adopts a dual-stream structure, targeting the extraction of tampering signs from both the spatial and DCT frequency domains. Within each stream, multi-scale units are employed to discern editing artifacts across varying receptive field sizes. Comprehensive comparative results indicate that our approach outperforms existing methods in the field of FAE localization, setting a new benchmark in performance. Additionally, when applied to the task of pinpointing facial image inpainting, our method demonstrated commendable results.