New attack reveals privacy risks in tabular foundation models

Researchers have identified significant privacy vulnerabilities in tabular foundation models, particularly within their attention layers. A new attack, AMIA, leverages transformer attention patterns to effectively perform Membership Inference Attacks (MIAs), demonstrating a notable increase in membership leakage compared to traditional methods. To address this, a novel defense mechanism inspired by k-anonymity principles has been developed, which reduces leakage by targeting high-risk queries without compromising model performance. The study also highlights that fine-tuning these models can further exacerbate privacy risks by amplifying memorization and exposing sensitive information through confidence shifts. AI

IMPACT Highlights potential privacy risks in widely used tabular foundation models, necessitating new defense strategies for sensitive data.

RANK_REASON Academic paper detailing a new attack and defense for AI model privacy. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AMIA
• Attention layers
• k-anonymity
• Membership Inference Attacks
• Tabular foundation models
• Tânia Carvalho Dr

• paper
• safety

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