New framework unifies membership inference attacks across generative models

Researchers have developed a unified framework for membership inference attacks (MIA) that can be applied across various generative model modalities, including text-to-text, text-to-image, and image-to-text. This new approach addresses the limitation of previous methods that treated each modality in isolation. The framework leverages the observation that a generative model's output distribution can approximate its training data distribution, enabling membership inference through likelihood ratio testing in a shared embedding space. Extensive experiments in black-box settings demonstrated superior performance compared to existing single-class optimized methods. AI

IMPACT Enhances understanding of privacy risks in generative models and provides new tools for auditing their training data.

RANK_REASON Academic paper published on arXiv detailing a new framework for membership inference attacks on generative models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• black-box setting
• fine-tuning
• Generative Models
• Image-to-text recognition for a sequence of images
• Membership Inference Attacks
• Partial Knowledge
• pre-training
• text-to-image model
• zero-knowledge proof

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →