Content Robust Image Generator Attribution

TMLR Paper8905 Authors

Abstract:

Image generator attribution aims to identify what generator produced an image, if any.
Prior work often focused on identifying new generators without requiring large amounts
of labeled samples by searching for shifts in image distributions. However, these shifts
may appear in other contests, such as a change in image content. Thus, an image may
be attributed to the wrong generator because its image content changed from what was
typically seen during training. To address this issue, we explore Content Robust imagE
generator attrIbuTion (CREdIT), where a model is evaluated on its ability to attribute an
image accurately even if the generators and/or image content is different than what was
seen during training. After a thorough analysis, we created a carefully crafted yet simple
baseline we refer to as FakesSense, which outperforms the state-of-the-art by 3-7%. This
illustrates a significant shortcoming in prior work, demonstrating a need for more complex
image generator attribution benchmarks like CREdIT.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Feng_Liu2

Submission Number: 8905