Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 17 papers Another 233 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 TRANSCOPE，用 CPU 硬件计数器推断训练集成员
Published Oct 6
Evaluated models BERT (4.4M-340M) , ViT (22M-307M) , Pythia (70M-2.8B) and others5 Abstract TRANSCOPE 揭示了分词器在嵌入表引发的硬件访问局部性差异，证实微架构计数器可突破恒定时间与置信度遮蔽防御。
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 提出用扰动蒸馏本地代理并离线过滤来提取黑盒模型凭据
Published Sep 29
Evaluated models DeepSeek-Coder-7B , LLaMA-3-8B , Qwen2.5-Coder-7B and others6 Abstract 论文提出扰动蒸馏与离线过滤两阶段框架，实现了 output-only 访问下延迟降低的黑盒大模型 API key 提取。
提出用邻居分数替代参考模型的单轮成员推断
Published Sep 29
Evaluated models 目标分类模型（架构与训练流程遵循 Aerni et al.） Abstract 邻居查询在单模型上恢复逐样本校准，合成邻居加早期检查点最接近 LiRA。本文研究 one-round 成员推断 ：只有一个目标模型、不能再训练参考模型时，如何仍做逐样本校准，并用于单次运行的 DP 审计。
提出用 API 推理摘要重建轨迹再蒸馏加强化学习，攻破蒸馏防御
Published Sep 29
Evaluated models Qwen2.5-0.5B , Qwen2.5-1.5B , Qwen2.5-3B and others12 Abstract 更现实的蒸馏攻击是蒸馏后再 RL。这篇工作修订蒸馏攻击的威胁模型：攻击者经合法 API 收集闭源推理数据，先蒸馏自有模型，再做强化学习。作者认为只在蒸馏后评估防御，会漏掉后续 RL，从而高估防御。
提出主题锚点投毒，放大众包微调后的专有指令提取
Published Sep 28
Evaluated models Qwen2.5-7B-Instruct , Qwen2.5-14B-Instruct , Llama-3.1-8B-Instruct and others4 Abstract 论文提出基于主题锚点的数据中毒攻击，证明低比例的中毒样本可在黑盒条件下隐蔽放大微调指令泄漏。
提出 Leaky，用新鲜轨迹的 token 对数概率差推断 OPD 训练记录成员身份
Published Sep 27
Evaluated models Qwen3-1.7B , Qwen3-4B , Qwen3.5-2B and others5 提出扩散潜空间地理扰动防御，抑制视觉语言模型泄露精确地理位置
Published Sep 18
Evaluated models GPT-4.1 , GPT-5 , Gemini 2.5 Pro and others5 提出 RAPID，离线把抗蒸馏扰动写入文生图解码器
Published Sep 15
Evaluated models PixArt , SSD , SDXL and others4 提出严格黑盒成员推断 WPMIA，用采样估计词级概率判断训练成员
Published Sep 8
Evaluated models OPT-6.7B , Pythia-6.9B , LLaMA-13B and others12 Abstract WPMIA 从续写估计词级似然并对比前缀，黑盒上多数优于所列基线。WPMIA 是面向严格黑盒的 LLM 预训练数据成员推断：攻击者只能看到文本续写。
提出针对微调 TTS 的黑盒成员推断攻击
Published Sep 2
Evaluated models CosyVoice2 , F5-TTS , XTTS-v2 Abstract 黑盒 TTS 成员推断以 recitation 加分层语音表示为分数，说话人级强于记录级，DP-SGD 可压到随机附近。作者给出一个针对微调文本转语音模型的黑盒成员推断框架，分别审计说话人是否进入微调集，以及某一条文本–音频记录是否进入微调集。微调 TTS 以合成文本和参考语音双条件生成，既有生成模型 MIA 的单条件查询和图像/文本上的逐点比较不能直接套用。作者用 scorable extent 与 memorization elicitation 比较五种查询，认为只有 recitation（用目标转写和干净目标语音要求模型复现）同时满足两条，并把它作为主查询。
利用加密思维块的跨模型重放，诱导弱模型逐字转写专有推理链
Published Aug 11
Evaluated models Claude Opus 4.8 , Claude Sonnet 4.6 , Claude Haiku 4.5 and others13 Abstract 论文揭示了客户端加密思维块跨模型互换漏洞，提出了低成本提取方法并给出防御方案。
Extract system instructions with encoding and structured requests, and harden them by rewriting with CoT
Published Apr 1
Evaluated models GPT-4.1-mini , GPT-3.5-turbo , Gemini-2.5-flash and others7 Abstract Encoded requests can elicit system instructions after a direct refusal.This work evaluates and hardens the confidentiality of LLM system instructions under encoded extraction. The authors are concerned with: when direct extraction has been rejected, whether the model still surrenders protected content when the same request is rewritten as a formatting or encoding task.
Proposing OPTILEAK, reconstructing others' prompts token-by-token via a KV cache side channel
Published Feb 24
Evaluated models Qwen-2.5-3B-Instruct , Qwen-2.5-7B-Instruct , Qwen-2.5-14B-Instruct and others4 Abstract OPTILEAK uses SFT and hard-token DPO to reduce the number of requests per token; the authors claim the risk of cache prompt leakage is higher than previously reported.
提出 Zero2Text，无需训练地跨域反演文本嵌入
Published Feb 2
Evaluated models GTR-Base , Qwen3-Embedding-0.6B , Text-Embedding-3-small and others4 Abstract 零训练在线对齐可在未知域、无泄漏对时从黑盒 embedding 还原句子。Zero2Text 是一种不训练 decoder、也不使用泄漏文本–向量对的文本 embedding 反演，针对 RAG 向量库在严格黑盒、目标域未知时的还原。
提出 JustAsk，用在线技能选择提取模型系统提示词
Published Jan 29
Evaluated models anthropic/claude-opus-4.5 , openai/gpt-5.2-codex , GPT-5.2 and others19 Attacks arXiv 2512.16059 · Dec 18, 2025 提出 ContextLeak，用金丝雀成员推断审计私有 ICL 泄漏
Published Dec 18, 2025
Evaluated models Llama3.3-70b , Qwen2.5-72b , GPT-4.1 and others6 Abstract ContextLeak 审计私有 ICL 的 worst-case 泄漏。ContextLeak 是一套用成员推断审计私有 ICL 的框架，目标是给 worst-case 泄漏一个经验下界，而不是只引用理论 ε。
You have reached the end