最近一年的 AI 安全论文
本站共收录 21 篇- VirusCascade：劫持 LLM 推荐智能体的协同反思机制
VirusCascade: Hijacking Collaborative Reflection in LLM-Powered Recommender Agents
- CSR：跨架构统一攻击概念擦除，FLUX 露骨内容攻击成功率达 50.47%
Concept Score Relearning: A Unified Cross-Architecture Attack on Concept Erasure
- 研究：提示注入可改变 Jev 决策模型的行动概率
Decision Hijacking: Prompt Injection Attacks on Jev's Typed Probabilistic Decisions
- SynChain：诱导计算机使用 Agent 自建攻击链并跨任务持久化
SynChain: Inducing Computer-Use Agent Systems to Construct Their Own Attack Chains
- 面向良性用途的对抗攻击：视觉内容全生命周期主动防护综述
Adversarial Attacks for Good: A Survey of Proactive Protection across the Visual Content Lifecycle
- 通过混合随机平滑实现异构扰动下的认证鲁棒性
Certified Robustness under Heterogeneous Perturbations via Hybrid Randomized Smoothing
- FraudBench：用于检测 AI 生成欺诈性退款证据的多模态基准
FraudBench: A Multimodal Benchmark for Detecting AI-Generated Fraudulent Refund Evidence
- 追踪拒答的动态：利用潜在拒答轨迹实现稳健的越狱检测
Tracing the Dynamics of Refusal: Exploiting Latent Refusal Trajectories for Robust Jailbreak Detection
- MPU：迈向大语言模型安全且隐私保护的知识遗忘
MPU: Towards Secure and Privacy-Preserving Knowledge Unlearning for Large Language Models
- AdapTools：针对智能体大语言模型的自适应工具型间接提示注入攻击
AdapTools: Adaptive Tool-based Indirect Prompt Injection Attacks on Agentic LLMs
- ICON：基于推理时纠正的 Agent 间接提示注入防御
ICON: Indirect Prompt Injection Defense for Agents based on Inference-Time Correction
- VENOMREC：针对多模态 LLM 推荐系统定向推广的跨模态交互投毒
VENOMREC: Cross-Modal Interactive Poisoning for Targeted Promotion in Multimodal LLM Recommender Systems
- 破坏层级推理：面向多模态推理模型地理隐私的对抗防护
Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models
- MultiPriv：视觉语言模型个体级隐私推理基准
MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models
- DualTAP：面向移动 MLLM Agent 的双任务对抗防护器
DualTAP: A Dual-Task Adversarial Protector for Mobile MLLM Agents
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。