UC Santa Barbara

Recent advances in image generation techniques have led to a proliferation of fake content online. This has created a need for adaptable, data driven methods to authenticate visual media, and prevent the spread of mis-information. In this defense, I will address two specific problem areas in detail.

First, the detection of GAN generated images. We demonstrate the efficacy of a novel forensic feature when combined with a conventional CNN architecture. Then, we investigate these in the adversarial setting, evaluating robustness to several levels of attack. These works represent some of the first attempts at adversarially robust GAN image detection.

The second area investigates methods for applying forensics to satellite images. Specifically, we look at the problems of metadata verification and image splicing detection. We demonstrate the power of incorporating satellite image metadata into the splicing detection problem, resulting in a joint model to accomplish both metadata and image verification with high accuracy.

The dissertation will conclude with a brief discussion of several other forensic areas, and potential future directions for the field.