最近一年的 AI 安全论文
本站共收录 76 篇- CHILLGuard：基于可扩展数据构建与模型感知偏好对齐的细粒度中文大模型安全护栏
CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment
- MemVenom：针对网页智能体多模态记忆的触发式投毒攻击
MemVenom: Triggered Poisoning of Multimodal Memories in Web Agents
- TS-LFO：通过两阶段潜在特征优化绕过扩散模型定制化中的版权保护
Bypassing Copyright Protection in Diffusion-based Customization via Two-Stage Latent Feature Optimization
- FlowErase-RL：将流匹配模型中的概念擦除重构为奖励优化
FlowErase-RL: Rethinking Concept Erasure as Reward Optimization in Flow Matching Models
- Mistletoe：针对大语言模型投机解码的隐蔽加速崩溃攻击
Mistletoe: Stealthy Acceleration-Collapse Attacks on Speculative Decoding
- 对齐印记（Alignment Imprint）：基于可证明偏好差异的零样本 AI 生成文本检测
Alignment Imprint: Zero-Shot AI-Generated Text Detection via Provable Preference Discrepancy
- C-ReD：基于真实提示词的综合性中文 AI 生成文本检测基准
C-ReD: A Comprehensive Chinese Benchmark for AI-Generated Text Detection Derived from Real-World Prompts
- 通过分层特征优化增强联邦学习中的梯度反演攻击
Enhancing Gradient Inversion Attacks in Federated Learning via Hierarchical Feature Optimization
- Weak-OOD 为何有效？进一步理解视觉语言模型的越狱
Why Does Weak-OOD Help? A Further Step Towards Understanding Jailbreaking VLMs
- JPRO：基于多智能体协作框架的自动化多模态越狱方法
JPRO: Automated Multimodal Jailbreaking via Multi-Agent Collaboration Framework
- CRDA：基于强化学习的自适应数据增强改进深度伪造检测
Improving Deepfake Detection with Reinforcement Learning-Based Adaptive Data Augmentation
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。