Double Face: Leveraging User Intelligence to Characterize and Recognize AI-synthesized Faces

Artificial Intelligence (AI) techniques have advanced to generate face images of nonexistent yet photorealistic persons. Despite positive applications, AI-synthesized faces have been increasingly abused to deceive users and manipulate opinions, such as AI-generated profile photos for fake accounts. Deception using generated realistic-appearing images raises severe trust and security concerns. So far, techniques to analyze and recognize AI-synthesized face images are limited, mainly relying on off-the-shelf classification methods or heuristics of researchers' individual perceptions. As a complement to existing analysis techniques, we develop a novel approach that leverages crowdsourcing annotations to analyze and defend against AI-synthesized face images. We aggregate and characterize AI-synthesis artifacts annotated by multiple users (instead of by individual researchers or automated systems). Our quantitative findings systematically identify where the synthesis artifacts are likely to be located and what characteristics the synthesis patterns have. We further incorporate user annotated regions into an attention learning approach to detect AI-synthesized faces. Our work sheds light on involving human factors to enhance defense against AI-synthesized face images.