Skip to content

Translation in progress Some entries still include text in the original language.

Oct 12 · Monday

All papers

Found 22 papers

Another 233papers do not have category tags yet and are not included in filtered results.

  1. AttacksarXiv 2609.35699

    论文:蒸馏防御在蒸馏后强化学习下易被攻破

    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
    Attacker
    Black-box
    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。

    Paper analysisRead analysis
  2. IDUnlearn-Bench:面向视觉语言模型的个体级遗忘基准

    提出 IDUnlearn-Bench,评测 VLM 个体级遗忘后的身份残留

    Published
    Evaluated models
    Qwen3-VL-2B, Qwen3-VL-4B, Qwen3-VL-8B and others6
    AbstractIDUnlearn-Bench 用四族探针显示:压低直接属性访问后,身份仍常能被反向查出、绑定或重构。

    IDUnlearn-Bench 评测 VLM 是否忘掉一个可跨模态识别的人,而不只是答不出某条监督事实。

    Paper analysisRead analysis
  3. AttacksarXiv 2609.27542

    Researchers use control token injection to delete gpt-oss-20b's reasoning chain and bypass chain-of-thought monitoring

    提出控制标记注入,清空工具 Agent 的思维链以绕过推理监控

    Published
    Setting
    AI Agent
    Evaluated models
    openai/gpt-oss-20b, google/gemma-4-26B-A4B-it, deepseek-ai/DeepSeek-R1-Distill-Qwen-7B and others4

    Abstract论文阐明控制标记注入可抑制 CoT 使监控失效并促成拒绝绕过,且工具执行受解析器宽容度影响。

    Paper analysisRead analysis
  4. AttacksarXiv 2609.15989

    Researchers propose plan injection attack that can bypass chain-of-thought monitoring

    提出 plan injection 攻击,植入良性伪装计划以绕过思维链监视器

    Published
    Attacker
    Gray-box
    Evaluated models
    Qwen2.5-Coder-7B-Instruct, Qwen3-8B, DeepSeek-R1-0528 and others6

    Abstract论文指出模型会将外部上下文中的伪装计划改写为自身推理,使思维链监视器失效甚至反向合理化恶意行为。

    Paper analysisRead analysis
  5. DefensesarXiv 2609.15671

    QPriv-VL:面向视觉语言模型的问题引导 token 裁剪隐私防御

    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
    Evaluated models
    LLaVA-1.5-7B, Llama-2
    AbstractQPriv-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。

    Paper analysisRead analysis
  6. SinoGlyphBench:中文字形混淆下语言模型审核的诊断基准

    提出 SinoGlyphBench,用配对字形混淆诊断中文审核稳定性

    Published
    Evaluated models
    GPT-5.4 mini, GPT-5.5, DeepSeek V4 Flash and others12
    AbstractSinoGlyphBench 显示中文字形混淆会同时增加有害漏报和误报,全文扰动最重。

    SinoGlyphBench是一个配对诊断基准,用来看内容审核在可还原的中文字形混淆下是否还保持原决定。

    Paper analysisRead analysis
  7. DefensesarXiv 2608.12876

    SPARED:用对抗编辑数据训练基于推理的 AI 生成图像检测器

    提出 SPARED,用对抗编辑数据训练基于推理的 AI 生成图像检测器

    Published
    Evaluated models
    Qwen3.5-9B
    AbstractSPARED 用防捷径的编辑对抗环训练 9B 推理检测器。

    SPARED 是一套用于可解释 AI 生成图像检测的攻防强化学习流程,报告模型是 Qwen3.5-9B 上的第三轮防御器。

    Paper analysisRead analysis
  8. AttacksarXiv 2608.09867

    Researchers steal proprietary LLM inference chains via cross-model compatibility of encrypted inference blocks

    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
    Attacker
    Black-box
    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。

    Paper analysisRead analysis
  9. DefensesarXiv 2607.14737

    GeoDetect:面向视觉语言预训练模型的几何对抗检测方法

    提出 GeoDetect 以几何分数检测视觉语言预训练模型的对抗样本

    Published
    Attacker
    White-box
    Evaluated models
    CLIPViT (ViT-B/16), CLIPCNN (ResNet-50), ALBEF and others4
    AbstractGeoDetect 用几何分数检测 VLP 对抗样本。

    GeoDetect 是面向视觉-语言预训练模型的对抗样本检测方法,用嵌入几何而不是任务 logits 区分干净样本与对抗样本。

    Paper analysisRead analysis
  10. AttacksarXiv 2605.30040

    研究揭示 LLM 按 token 计费可被服务商系统性虚报

    构造可通过计费审计的 token 虚报,使服务商系统性超收

    Published
    Evaluated models
    PALACE domain-adapted auditor
    Abstract不诚实提供方可在三种 token 审计上虚报并通过检查,作者认为需改用提供方不控制的证据。

    不透明 LLM 的按 token 计费里,现有审计核验的是提供方可控的证据,不诚实提供方可以虚报并通过检查。。

    Paper analysisRead analysis
  11. AttacksarXiv 2605.21541

    研究者提出 FRA-Attack 频域正则化方法,提升对闭源 MLLM 的迁移攻击效果

    提出 FRA-Attack,用频域对齐与梯度正则做可迁移定向视觉攻击

    Published
    Attacker
    Black-box
    Evaluated models
    GPT-5.2, GPT-5.4, Claude-Opus-4.6 and others15
    AbstractFRA-Attack 以频域对齐和梯度低通做可迁移定向攻击,闭源面板上作者称整体更强,Gemma 除外。

    FRA-Attack 是面向闭源 MLLM 的迁移式定向图像攻击。作者要解决的是:空间域局部对齐重复全局低频,代理梯度的高频成分又把更新拉向代理自身。

    Paper analysisRead analysis
  12. AttacksarXiv 2604.03117

    Unveiling Physical-World Semantic Vulnerabilities: Universal Adversarial Patches for Infrared Vision-Language Models

    Proposes the universal curved mesh patch UCGP, causing cross-task semantic bias in infrared vision-language models.

    Published
    Attacker
    Gray-box
    Evaluated models
    OpenAI CLIP ViT-L/14, OpenCLIP ViT-B/16, Meta-CLIP ViT-L/14 and others10
    AbstractUCGP uses one shared curved-grid patch to simultaneously lower the classification and open-ended outputs of IR-VLM.

    UCGP is a universal physical adversarial patch for infrared vision-language models. The question is: can a patch that can be attached and reused across samples cause stable semantic deviations in the infrared visual evidence shared by classification, description, and VQA. The patch is constrained to a low-frequency curved mesh.

    Paper analysisRead analysis
  13. DefensesarXiv 2602.24210

    研究者提出 Staged Decoding 与 SFT 数据集,减少大推理模型推理轨迹中的隐私泄露

    提出 Staged Decoding 与指令跟随训练,减少推理轨迹隐私泄露

    Published
    Evaluated models
    Qwen 3 1.7B, Qwen 3 4B, Qwen 3 8B and others6
    Abstract用一般指令跟随训练加分段解码,可降低轨迹隐私泄露,但数学效用下降。

    本文把大推理模型的隐私问题做成轨迹可控性:隐私指令是指令,轨迹不跟随就会把敏感上下文写进可被提示注入导出的文本。

    Paper analysisRead analysis
  14. DefensesarXiv 2602.18252

    Adversarial robustness of discrete image tokenizers

    提出无监督对抗微调,加固离散图像 tokenizer

    Published
    Evaluated models
    UniTok, UniTok-MLLM, TiTok-BL128 and others7
    Abstract无监督嵌入攻击暴露离散 tokenizer 的弱点,只微调 encoder 即可提升多任务鲁棒准确率。

    作者研究离散图像 tokenizer 的对抗鲁棒性,给出与任务无关的嵌入攻击,以及只微调 encoder 的无监督对抗训练。

    Paper analysisRead analysis
  15. DefensesarXiv 2602.12461

    Semantic-aware adversarial fine-tuning for CLIP

    Propose SAFT, which uses semantic description ensembles to construct adversarial examples and fine-tune the image encoder

    Published
    Evaluated models
    ViT-B/32, CLIP-L/14
    AbstractSAFT fine-tunes the CLIP image encoder with ensemble adversarial examples from refined class descriptions to improve the trade-off between zero-shot accuracy and robustness.

    SAFT is a method for adversarially fine-tuning the CLIP image encoder to improve the accuracy–robustness trade-off in zero-shot classification.

    Paper analysisRead analysis
  16. AttacksarXiv 2602.09431

    Grounding-driven attack: Improving the transferability of encoder-based adversarial attacks to large vision-language models

    Propose GDA, which aligns image adversarial perturbations with text-grounded evidence and transfers attacks to vision-language models

    Published
    Evaluated models
    LLaVA, Qwen2.5-VL, InternVL3 and others10
    AbstractGDA aligns perturbations with text-grounded evidence, and the authors say this improves zero-query black-box transferability on heterogeneous LVLMs.

    GDA is a zero-query black-box encoder transfer attack for heterogeneous LVLMs.

    Paper analysisRead analysis
  17. AttacksarXiv 2601.21531

    Adversarial robustness of large vision-language models under visual token compression

    提出 CAGE 攻击,在视觉 token 压缩下用图像扰动使预测偏离

    Published
    Attacker
    Gray-box
    Evaluated models
    LLaVA, Qwen2.5-VL, InternVL3-8B and others5
    AbstractCAGE 对齐压缩瓶颈后,压缩 LVLM 的稳健准确率低于全 token 编码器攻击。

    CAGE 研究的是:LVLM 在推理时剪枝或合并视觉 token 之后,对抗稳健性该怎么估。

    Paper analysisRead analysis