Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 16 papers Another 232 papers do not have category tags yet and are not included in filtered results.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Annotated first” shows papers with tags or published paper analyses first. Each group is ordered by original publication date, newest first, with undated papers last. “Newest first” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and paper analyses are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
提出 DP-MixMin,私有学习公开数据混合并做差分隐私微调
Published Oct 6
Evaluated models ResNet-18 , GPT-2 Medium Abstract DP-MixMin 私下学公开混合再 DP 微调,胸片 macro AUC 与邮件困惑度优于均匀混合。DP-MixMin 是一条端到端的差分隐私训练流程:先私下决定多个公开数据集的预训练混合,再预训练并在敏感数据上 DP 微调。作者称这是首个这样的流程。
提出 Persona Guardrail,强制客服智能体守住功能边界
Published Oct 2
Evaluated models Qwen3-30B-A3B , Qwen3-VL-30B , Qwen3-8B and others5 Abstract Persona Guardrail 用功能边界做同步输入输出校验,并在 PAGE 与生产延迟预算下给出评测。Persona Guardrail 是客服向智能体的同步运行时防御:先写明功能边界,再决定放行或拦截,而不是只检测某一攻击类。
提出 COVER,在语言预算下选择源语言实现跨语言知识遗忘
Published Oct 1
Evaluated models Tiny Aya Global , Qwen3-4B-Instruct-2507 , Llama-3.2-3B-Instruct Abstract COVER 在语言预算下选源语言,以降低未监督语言上相对 oracle 的残余访问。语言预算下的多语言遗忘 :遗忘监督只放在 K 种源语言上,目标是降低其余语言对同一事实的残余访问,而不是把每种评测语言都译成监督数据。
提出 Whiteout,用伪装样本微调覆盖个人敏感信息关联
Published Oct 1
Evaluated models Llama 3.2 , Phi-3-mini , Gemma 3 and others7 提出 Intervention Score 与 DIR-R,按干预效果选择并修订遗忘参数支持
Published Sep 29
Evaluated models Llama 3 8B , Gemma 2 2B , Qwen2.5-7B and others4 Abstract 作者把局部化遗忘拆成初始干预选择和检查点修订,并在三个设置里分别给证据和边界。Storage Is Not Strategy 研究局部化 LLM 遗忘里两个常被混在一起的决定:一开始该更新哪些参数,以及优化改变模型之后这个选择要不要改。
提出 TULIP,按输入定位形成层并去除目标 token 对齐以实现定向遗忘
Published Sep 28
Evaluated models Llama-3.1-8B , Llama-3.2-3B , Llama-3.2-1B and others7 Abstract TULIP 在形成结束的层去掉遗忘目标对齐,多数设置下 FQ 更高,4-bit 下 FQ 下降。TULIP 是一种逐输入选择层的表示级遗忘方法,目标是去掉遗忘答案的形成、留下读出。输出级方法可能只压低特定回答的似然。RMU、LUNAR 等则对全部遗忘样本共用一层。
提出时间步感知目标条件化,改善差分隐私能源时序插补效用
Published Sep 21
Evaluated models DiT-style Transformer Abstract 固定阈值 DP-SGD 下,v-prediction 加 ᾱt 加权降低能源时序插补误差,并降低后期低 SNR 的裁剪前梯度。作者研究固定阈值 DP-SGD 训练的扩散模型,如何在余弦调度的后期低 SNR 时间步避免 ε-prediction 产生易被裁剪的大梯度,从而做能源时序缺失值恢复。
提出 AUDIT PLAN 先计划后作答,并用 FAITHGATE 绑定答案奖励与计划正确性
Published Sep 17
Evaluated models Qwen2.5-1.5B-Instruct , Qwen2.5-3B-Instruct , Qwen-3-4B-Instruct and others4 Abstract 先承诺可机检安全计划,再让计划正确性决定能否拿到答案奖励。AUDIT PLAN 把安全对齐从“只优化最终答案”改成“先提交隐藏的结构化安全计划,再按该计划作答”,并用 FAITHGATE 把高答案奖励限定在计划正确时。
提出 QPriv-VL,用问题引导剪枝降低分布式视觉问答的隐私泄露
Published Sep 14
Evaluated models LLaVA-1.5-7B , Llama-2 Abstract QPriv-VL 在切层前按问题相关度与 DINOv2 敏感度剪视觉 token,作者称能在约四成 token 下兼顾准确率与四类攻击。QPriv-VL 把问题引导的视觉 token 剪枝做成分布式 VQA 的客户端隐私防御,而不是只做集中式推理加速。
提出 Adaptive Diffusion Freezing,降低扩散模型的成员推理风险
Published Sep 8 Proposes SHAQ, which replaces document embeddings with shadow query embeddings to defend against embedding inversion.
Published Sep 4
Evaluated models GTR-T5 encoder , DPR Abstract SHAQ replaces document embeddings with shadow query embeddings.SHAQ is a defense against EIA for cloud vector databases: instead of directly modifying document embeddings, it replaces them with shadow query embeddings.
提出 Skill-as-API,隔离多智能体协作中的 skill 正文与系统提示词
Published Sep 1 Abstract 协议把 skill 正文留在所有者侧,四层限制发现内容、权限期限和注入拼接。Skill-as-API 是面向智能体软件工程的协调协议,用来在跨组织调用时把 skill 的代码和系统提示词留在所有者进程。作者针对三条威胁:经通道或命名空间抽取专有方法,用注入让被调用模型输出系统提示词,以及协作结束后权限继续保留。
提出 LeakGauge,用行为后缀的 prefill 对数概率检测上下文泄露
Published Aug 18
Evaluated models Gemma-4-E4B-it , Gemma-4-31B-it , Qwen2.5-14B-Instruct and others11 Propose soft prompt defense AREA, re-anchoring attention to mitigate system prompt leakage
Published Jun 17
Evaluated models Llama-2-7B-chat-hf , Llama-3.1-8B-Instruct , Mistral-7B-Instruct-v0.3 and others8 Abstract The authors measure commercial prompt leakage, propose attention drift, and use AREA to strike a trade-off between usability and leakage resistance.The authors study system prompt leakage in real LLM applications and propose AREA, a mitigation method that does not modify weights.
提出零知识审计协议,核验多租户 RAG 的合谋差分隐私
Published May 19 Abstract 揭示多租户 RAG 合谋泄露按根号 k 放大,提出带零知识证明的无泄露审计协议。论文针对多租户 RAG 服务在多账户协同攻击下隐私边界失效的问题展开研究。
提出 SeqWM,把水印写入智能体动作转移并用随机密钥检测
Published May 11
Evaluated models LLaMA-3.2 3B Instruct , Gemma 4 E4B-it , Qwen3-4B Instruct Abstract SeqWM 以条件转移嵌行为水印。SeqWM 是给 LLM 智能体动作序列加水印的框架,用来在只看到动作时判断轨迹是否来自持有密钥的带水印源。
You have reached the end