paper

Privacy Vulnerabilities of Attention Layers in Tabular Foundation Models and Protection of High-Risk Queries

paperactiveprovisionalprivacy-vulnerabilities-of-attention-layers-in-tabular-foundation-models-and-protection-of-high-risk-queries-898e5eb4·1 events·first seen 3d ago

Aliases: Privacy Vulnerabilities of Attention Layers in Tabular Foundation Models and Protection of High-Risk Queries

Co-occurring entities

Membership Inference Attack AMIA

More like this (12)

Your Privacy My Cloak: Backdoor Attacks on Differentially Private Federated Learning representation-level sensitive information leakage Tabular Foundation Models (TFMs)Privacy Inference Attack Policy Vulnerability Testing Clinically Grounded Privacy Evaluation of Medical LMs Security and Privacy Prompts in the Wild: What Users Ask LLMs and How LLMs Respond Grouped-Query Attention Speculative Tool Privacy Contracts Predictability as a Fine-Grained Measure for Privacy Privacy-Preserving RAG via Multi-Agent Semantic Rewriting Beyond Runtime Enforcement: Shield Synthesis as Defensibility Analysis for Adversarial Networks

Recent events (1)

5arXiv · cs.AI·3d ago·source ↗

AMIA: Attention-based membership inference attacks on tabular foundation models with k-anonymity defense

Researchers demonstrate that tabular foundation models using in-context learning are vulnerable to membership inference attacks (MIAs) via attention mechanism leakage, even when pre-trained on synthetic data. They introduce AMIA, a shadow-model-free attack exploiting transformer attention concentration patterns, achieving a 7.7% average gain over confidence-based attacks. A k-anonymity-inspired inference-time defense reduces membership leakage by 50% against AMIA and 25% against confidence-based attacks with only 3.9% performance degradation. The paper also shows fine-tuning amplifies memorization risk through confidence shifts.

Evaluation and Benchmarking AI Safety Research Membership Inference Attack Privacy Vulnerabilities of Attention Layers in Tabular Foundation Models and Protection of High-Risk Queries AMIA