顶会安全论文

ICLR、ICML、NeurIPS、ACL 和安全四大会（IEEE S&P、USENIX Security、ACM CCS、NDSS）录用的 AI 安全论文，按年份和研究方向整理。以会议的安全 track 为主，其他 track 里内容明确是安全方向的也收，并标出来。

共 45 篇

越狱与攻击

4

EvilHarmony：针对黑盒语音识别系统的隐蔽对抗攻击EvilHarmony: Stealthy Adversarial Attacks against Black-box Speech Recognition Systems
非 AI 分会场Xuejing Yuan, Jiangshan Zhang, Feng Guo 等会议页面
模糊测试结合LLM智能体：自动高效越狱文本到图像生成模型Fuzz-Testing Meets LLM-Based Agents: An Automated and Efficient Framework for Jailbreaking Text-To-Image Generation Models
Yingkai Dong, Xiangtao Meng, Ning Yu 等会议页面
Modifier Unlocked：通过提示词越狱文本到图像模型Modifier Unlocked: Jailbreaking Text-to-Image Models Through Prompts
Shuofeng Liu, Mengyao Ma, Minhui Xue 等会议页面
UnMarker：针对防御性图像水印的通用攻击UnMarker: A Universal Attack on Defensive Image Watermarking
Andre Kassis, Urs Hengartner会议页面

提示注入

2

DataSentinel：基于博弈论的提示注入攻击检测DataSentinel: A Game-Theoretic Detection of Prompt Injection Attacks
Yupei Liu, Yuqi Jia, Jinyuan Jia 等会议页面
Fun-tuning：专有LLM对基于优化的提示注入攻击的脆弱性Fun-tuning: Characterizing the Vulnerability of Proprietary LLMs to Optimization-based Prompt Injection Attacks via the Fine-Tuning Interface
Andrey Labunets, Nishit V. Pandya, Ashish Hooda 等会议页面

防御与护栏

12

实际深度学习无线系统中基于 ViT 的对抗鲁棒自动调制识别Adversarial Robust ViT-based Automatic Modulation Recognition in Practical Deep Learning-based Wireless Systems
非 AI 分会场Gen Li, Chun-Chih Lin, Xiaonan Zhang 等会议页面
EveGuard：用音频对抗扰动抵御基于振动的侧信道窃听EveGuard: Defeating Vibration-based Side-Channel Eavesdropping with Audio Adversarial Perturbations
非 AI 分会场Jung-Woo Chang, Ke Sun, David Xia 等会议页面
探究亲子视角下的生成式 AI 安全：担忧、缓解策略与设计启示Exploring Parent-Child Perspectives on Safety in Generative AI: Concerns, Mitigation Strategies, and Design Implications
非 AI 分会场Yaman Yu, Tanusree Sharma, Melinda Hu 等会议页面
以火攻火：用模式随机化防御补丁对抗对抗性补丁攻击Fight Fire with Fire: Combating Adversarial Patch Attacks using Pattern-randomized Defensive Patches
Jianan Feng, Jiachun Li, Changqing Miao 等会议页面
GuardAIn：保护异构 NPU 上的新兴生成式 AI 工作负载GuardAIn: Protecting Emerging Generative AI Workloads on Heterogeneous NPU
非 AI 分会场Aritra Dhar, Clément Thorens, Lara Magdalena Lazier 等会议页面
协作机器学习中鲁棒性与学习之间的冲突On the Conflict between Robustness and Learning in Collaborative Machine Learning
Mathilde Raynal, Carmela Troncoso会议页面
查询溯源分析：抵御基于查询的黑盒攻击的高效稳健防御Query Provenance Analysis: Efficient and Robust Defense against Query-based Black-box Attacks
Shaofei Li, Ziqi Zhang, Haomin Jia 等会议页面
SoK：AI 生成内容的水印技术SoK: Watermarking for AI-Generated Content
Xuandong Zhao, Sam Gunn, Miranda Christ 等会议页面
用大语言模型辅助人工评审员检测有害内容Supporting Human Raters with the Detection of Harmful Content using Large Language Models
非 AI 分会场Kurt Thomas, Patrick Kelley, David Tao 等会议页面
TSQP：保护边缘设备上量化神经网络的实时推理TSQP: Safeguarding Real-Time Inference for Quantization Neural Networks on Edge Devices
Yu Sun, Gaojian Xiong, Jianhua Liu 等会议页面
可验证的提升树集成模型Verifiable Boosted Tree Ensembles
Stefano Calzavara, Lorenzo Cazzaro, Claudio Lucchese 等会议页面
面向众多自适应用户的语言模型水印Watermarking Language Models for Many Adaptive Users
Aloni Cohen, Alexander Hoover, Gabe Schoenbach会议页面

对齐

1

缓解 LLM 微调中丢失对齐的担忧Alleviating the Fear of Losing Alignment in LLM Fine-tuning
Kang Yang, Guanhong Tao, Xun Chen 等会议页面

奖励作弊

1

奖励模型学习中的偏好投毒攻击Preference Poisoning Attacks on Reward Model Learning
Junlin Wu, Jiongxiao Wang, Chaowei Xiao 等会议页面

后门与投毒

6

从第一性原理研究架构神经后门Architectural Neural Backdoors from First Principles
Harry Langford, Ilia Shumailov, Yiren Zhao 等会议页面
BAIT：通过反演攻击目标扫描大语言模型后门BAIT: Large Language Model Backdoor Scanning by Inverting Attack Target
Guangyu Shen, Siyuan Cheng, Zhuo Zhang 等会议页面
并非所有边都同样鲁棒：评估基于排序的联邦学习的鲁棒性Not All Edges are Equally Robust: Evaluating the Robustness of Ranking-Based Federated Learning
Zirui Gong, Yanjun Zhang, Leo Yu Zhang 等会议页面
PEFTGuard：检测针对参数高效微调的后门攻击PEFTGuard: Detecting Backdoor Attacks Against Parameter-Efficient Fine-Tuning
Zhen Sun, Tianshuo Cong, Yule Liu 等会议页面
聚类联邦学习中有限拜占庭客户端的实用投毒攻击Practical Poisoning Attacks with Limited Byzantine Clients in Clustered Federated Learning
Viet Vo, Mengyao Ma, Guangdong Bai 等会议页面
安全迁移学习：抵御预训练编码器与下游数据集中的后门Secure Transfer Learning: Training Clean Model Against Backdoor in Pre-Trained Encoder and Downstream Dataset
Yechao Zhang, Yuxuan Zhou, Tianyu Li 等会议页面

隐私与数据泄露

14

CipherSteal：利用密文侧信道从 TEE 保护的神经网络中窃取输入数据CipherSteal: Stealing Input Data from TEE-Shielded Neural Networks with Ciphertext Side Channels
非 AI 分会场Yuanyuan Yuan, Zhibo Liu, Sen Deng 等会议页面
Codebreaker：代码语言模型的动态提取攻击Codebreaker: Dynamic Extraction Attacks on Code Language Models
非 AI 分会场Changzhou Han, Zehang Deng, Wanlun Ma 等会议页面
Comet：通过预测激活稀疏性加速大语言模型的私有推理Comet: Accelerating Private Inference for Large Language Model by Predicting Activation Sparsity
非 AI 分会场Guang Yan, Yuhui Zhang, Zimu Guo 等会议页面
边缘遗忘并非“在边缘”：资源受限设备上的自适应精确遗忘系统Edge Unlearning is Not "on Edge"! An Adaptive Exact Unlearning System on Resource-Constrained Devices
Xiaoyu Xia, Ziqi Wang, Ruoxi Sun 等会议页面
GRID：保护训练图免受 GNN 模型的链接窃取攻击GRID: Protecting Training Graph from Link Stealing Attacks on GNN Models
Jiadong Lou, Xu Yuan, Rui Zhang 等会议页面
HARMONYCLOAK：让音乐对生成式AI不可学习HARMONYCLOAK: Making Music Unlearnable for Generative AI
Syed Irfan Ali Meerza, Lichao Sun, Jian Liu会议页面
面向受保护属性的隐私机器学习Machine Learning with Privacy for Protected Attributes
Saeed Mahloujifar, Chuan Guo, Edward Suh 等会议页面
真实环境提示词窃取攻击的有效性On the Effectiveness of Prompt Stealing Attacks on In-The-Wild Prompts
Yicong Tan, Xinyue Shen, Yun Shen 等会议页面
普及压过担忧？中国用户对基于 LLM 的医疗咨询的隐私意识与期望Prevalence Overshadows Concerns? Understanding Chinese Users' Privacy Awareness and Expectations Towards LLM-based Healthcare Consultation
非 AI 分会场Zhihuang Liu, Ling Hu, Tongqing Zhou 等会议页面
针对大语言模型协同推理的提示词反演攻击Prompt Inversion Attack against Collaborative Inference of Large Language Models
Wenjie Qu, Yuguang Zhou, Yongji Wu 等会议页面
操纵基础模型：通过预训练操控实现更强的成员推断攻击Rigging the Foundation: Manipulating Pre-training for Advanced Membership Inference Attacks
Zihao Wang, Rui Zhu, Zhikun Zhang 等会议页面
SoK：机器学习系统中的数据集版权审计SoK: Dataset Copyright Auditing in Machine Learning Systems
Linkang Du, Xuanru Zhou, Min Chen 等会议页面
可靠验证个性化文生图扩散模型中的未授权数据使用Towards Reliable Verification of Unauthorized Data Usage in Personalized Text-to-Image Diffusion Models
Boheng Li, Yanhao Wei, Yankai Fu 等会议页面
理解用户对对话式AI平台的安全与隐私担忧及态度Understanding Users' Security and Privacy Concerns and Attitudes Towards Conversational AI Platforms
Mutahar Ali, Arjun Arunasalam, Habiba Farrukh会议页面

供应链安全

2

我的模型对你来说是恶意软件：滥用TensorFlow API将AI模型转化为恶意软件My Model is Malware to You: Transforming AI Models into Malware by Abusing TensorFlow APIs
非 AI 分会场Ruofan Zhu, Ganhao Chen, Wenbo Shen 等会议页面
LLM应用商店的安全性分析On the (In)Security of LLM App Stores
Xinyi Hou, Yanjie Zhao, Haoyu Wang会议页面

危险能力

1

GPTracker：大规模测量GPT的滥用情况GPTracker: A Large-Scale Measurement of Misused GPTs
Xinyue Shen, Yun Shen, Michael Backes 等会议页面

AI 与网络攻防

2

声东击西：用Flashboom攻击蒙蔽基于LLM的代码审计器Make a Feint to the East While Attacking in the West: Blinding LLM-Based Code Auditors with Flashboom Attacks
非 AI 分会场Xiao Li, Yue Li, Hao Wu 等会议页面
SV-TrustEval-C：评估大语言模型在源码漏洞分析中的结构与语义推理SV-TrustEval-C: Evaluating Structure and Semantic Reasoning in Large Language Models for Source Code Vulnerability Analysis
非 AI 分会场Yansong Li, Paula Branco, Alexander M. Hoole 等会议页面