两篇新研究论文探讨了大型语言模型（LLM）的隐私漏洞。一篇论文介绍了一个数据集和评估框架，用于识别多模态LLM中的隐私风险，并强调了这些模型如何泄露图像和内存中的敏感信息。另一篇论文对适配LLM的差分隐私（DP）有效性进行了基准测试，发现数据分布变化显著影响隐私风险，并且像LoRA这样的参数高效微调方法能为分布外数据提供更好的保护。 AI
影响 强调了LLM隐私方面存在的关键漏洞，敦促开发人员为多模态和适配型模型实施强大的安全措施。
排序理由 两篇在arXiv上发表的学术论文，详细介绍了LLM的隐私风险和缓解策略。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2606.09125v1 Announce Type: cross Abstract: Privacy risks in text-only Large Language Models (LLMs) are well studied, particularly their tendency to memorize and leak sensitive information. However, Multi-modal Large Language Models (MLLMs), which process both text and imag…
arXiv:2606.09401v1 Announce Type: new Abstract: Recent work has applied differential privacy (DP) to adapt large language models (LLMs) for sensitive applications, offering theoretical guarantees. However, its practical effectiveness remains unclear, partly due to LLM pretraining…
Recent work has applied differential privacy (DP) to adapt large language models (LLMs) for sensitive applications, offering theoretical guarantees. However, its practical effectiveness remains unclear, partly due to LLM pretraining, where overlaps and interdependencies with adap…