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

2026-06-18 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

A new black-box behavioral probe, NAMESAKES, has been introduced to distinguish identity memorization from fabrication in Text-to-Image (T2I) models. T2I models can generate realistic likenesses from names, posing privacy risks, but existing detection methods require ground-truth photos, training data access, or white-box model internals. The NAMESAKES probe operates without these requirements, offering a fully black-box solution. To benchmark this task, the researchers created the NAMESAKES dataset, comprising over one thousand names and faces of public figures across various fame levels, alongside perturbed, less famous names. Experiments on leading T2I models, published on 2026-06-18, demonstrate that the probe accurately predicts identity memorization and differentiates between memorized and unrecognized names, providing insights into variations across model families.

Key takeaway

For AI Security Engineers evaluating Text-to-Image models for deployment, this black-box probe offers a critical tool to assess identity memorization risks. You can now identify models that generate realistic likenesses from names without needing internal access or ground-truth photos. This allows you to proactively mitigate privacy concerns and ensure compliance by selecting models less prone to memorizing specific public figures, enhancing responsible AI development.

Key insights

A black-box probe can reliably detect identity memorization in text-to-image models without needing ground truth or training data access.

Principles

• T2I models memorize specific identities from training data.
• Black-box probes can infer model memorization states.
• Fame level influences identity memorization.

Method

The NAMESAKES method uses a black-box behavioral probe to distinguish memorized from fabricated identities in T2I models. It benchmarks this using a dataset of public figures' names and perturbed names.

In practice

• Identify T2I models with privacy risks.
• Benchmark T2I models for memorization.
• Inform T2I training data curation.

Topics

• Text-to-Image Models
• Identity Memorization
• Privacy Concerns
• Black-box Probing
• NAMESAKES Dataset
• AI Security

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, AI Security Engineer, AI Ethicist

Related on AIssential

• MIT scientists investigate memorization risk in the age of clinical AI — AI models trained on EHRs can memorize patient data, necessitating rigorous testing to prevent privacy breaches.
• Phantoms and Disclosures: a Causal Framework for Auditing Synthetic Data — The framework distinguishes true from phantom data disclosures in synthetic data using statistical testing, requiring no model access.
• MRMMIA: Membership Inference Attacks on Memory in Chat Agents — Chat agent memory is vulnerable to membership inference attacks, which MRMMIA effectively demonstrates across various access levels.
• YouTube is expanding its AI deepfake detection tool to all adult users — YouTube's AI deepfake detection tool for adult users raises significant biometric data privacy concerns.
• PromptPrint: Behavioral Biometrics Through Natural Language Prompting in LLMs — User identity can be reliably inferred from LLM prompts based on surface-level lexical patterns.
• Detecting Data Contamination in Large Language Models — Black-box MIAs are currently ineffective at reliably detecting data contamination in LLM training corpora.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.