NAMESAKES: Probing Identity Memorization in Text-to-Image Models
Researchers have developed a new black-box method to detect if text-to-image models have memorized specific individuals' identities. This probe does not require ground-truth photos or access to the model's training data. To test their method, they created the NAMESAKES dataset, which includes over a thousand names and faces of public figures, along with less famous names. Experiments on current text-to-image models demonstrated that the probe can effectively predict identity memorization and differentiate between memorized and unrecognized individuals. AI
IMPACT This research could lead to better privacy controls for generative AI models, impacting how user data is handled and protected.