RI: Small: Understanding and Advancing the Generalization Capabilities of Fake Image Detectors

This project develops an integrated research, education, and outreach program to provide a foundation for understanding and advancing the generalization capabilities of fake (AI-generated) image detectors. With the rise and maturity of artificial intelligence (AI) generative models, fake images are proliferating in the digital world at an unprecedented rate. Fake images can cause harm in a variety of ways by deceiving, manipulating, and misinforming individuals and society at large. For example, they can be used for blackmail, disinformation, or financial fraud, etc. To make matters worse, there is no longer a single source of fake images. Synthesized images could take the form of realistic human faces generated using one type of fake image generator or they could take the form of complex scenes generated using another type of fake image generator. One can be almost certain that there will be more methods developed for generating fake images coming in the future. To combat the potential harm caused by fake images, this project aims to advance technology in image forensics, and build a deep understanding as to how AI-generated images differ from real images. In addition to scientific impact, this project performs complementary educational and outreach activities that engage students in research and STEM. This research is to provide a foundation for gaining a deeper understanding of, and for advancing fundamental research in, detecting AI-generated images. In particular, it will focus on the generalization properties of fake image detectors. Specifically, it will investigate two main thrusts: (Thrust I) Understanding what makes fake AI-generated images fake, including the role of generative models vs. data, a novel benchmark and toolbox for universal fake image detection, and understanding the features that a fake detector focuses on. In Thrust II, the project will advance universal fake AI-generated image detection, including a novel frequency masking strategy, a few-shot adaptation approach for learning with only a few training examples, and extensions to video to account for spatio-temporal aspects of realism. Both thrusts will be studied in the context of creating generalizable (universal) fake image detection algorithms. The investigators' wealth of experience in generative models and understanding model robustness, as well as initial work in this space, make them well-positioned to formulate and solve the relevant challenges, and lays the groundwork for the project. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.