NAMESAKES: Probing Identity Memorization in Text-to-Image Models

2026-06-18

AI Summary

This work addresses the risk that text-to-image models may inadvertently leak the identities of real individuals when generating portraits associated with personal names. Existing approaches struggle to distinguish whether such outputs stem from memorization or random generation, particularly in the absence of reference images or access to training data. To tackle this challenge, we propose the first fully black-box behavioral probing method that effectively differentiates identity memorization from stochastic generation without requiring white-box model access or prior knowledge of real photographs. We introduce NAMESAKES, a large-scale benchmark dataset comprising individuals of varying prominence and their perturbed name variants. Extensive cross-model-family experiments demonstrate that our probe reliably predicts identity memorization, accurately identifies names retained by the model, and reveals systematic differences in this behavior across distinct model families.

Abstract

Text-to-image (T2I) models generate realistic likenesses of some individuals when prompted with their names, raising privacy concerns. However, distinguishing whether a generated face is memorized or fabricated currently requires ground-truth photos, access to training data, or white-box access to model internals, limiting applicability. We introduce a fully black-box behavioral probe that distinguishes between these regimes while requiring no reference photos or prior knowledge of training data. To benchmark this task, we present the NAMESAKES dataset of over one thousand names and faces of public figures spanning a wide range of fame levels, along with perturbed, less famous names. Experiments on state-of-the-art T2I models show that our probe substantially predicts identity memorization and separates memorized from unrecognized names, with further insights into differences across model families.

Problem

• text-to-image models
• identity memorization
• privacy concerns
• black-box probing
• face generation

Innovation

• black-box probing
• identity memorization
• text-to-image models
• privacy
• NAMESAKES dataset