#membership inference

3 stories · 1 in 7 days · 2 in 30 days

LoRA 1 Membership Inference Attacks 1 OP-Print 1 SWAG 1 VLM 1 self-distillation 1 distillation 1 privacy 1

Stories, products and papers tagged "membership inference", newest first.

Nov 10

Papers Builders10:26

Chinese content用贝叶斯 LoRA 后验替代影子模型，让成员推断攻击在大型 VLM 上可行

做隐私审计或模型安全研究的可以看看这篇：训练一次 LoRA 就能生成虚拟影子模型，省掉传统 MIA 的模型集成开销，实测在 Document VQA 和 Medical VQA 上比标准攻击更强。

#membership inference #LoRA #SWAG #VLM

arXiv cs.LG@Khanh Nguyen et al. (6)Source

Oct 23

Papers Builders11:28

Chinese contentOP-Print：针对在线自蒸馏模型的成员推断攻击研究

一篇隐私方向的新论文：自蒸馏模型原来会留下可被攻击的训练数据痕迹，OP-Print 用残差信号和重启稳定性就能把成员识别出来，基线全都失效。

#membership inference #self-distillation #OP-Print #privacy

arXiv cs.LG@Qi Li, Xinchao WangSource

Sep 11

Papers BuildersSelected11:22

Predicting Privacy Leakage from Weight Spectral Density

This research provides a new method for evaluating privacy risks in machine learning models, which is important for developers and researchers working on AI systems.

#Membership Inference Attacks #privacy leaks #membership inference

arXiv cs.LG@Richard J. Preen, Jim SmithSource