大模型安全LLM Security
LLM & MULTIMODAL SECURITY
研究越狱攻击与防御、生成过程中的风险识别，以及大语言模型与多模态模型的可用性安全。Jailbreak attacks and defenses, risk detection during generation, and the availability of language and multimodal models.
相关成果Related work ↗TRUSTWORTHY AI · SECURE SYSTEMS
面向真实世界的
可信人工智能。Building AI
worthy of trust.
我是刘小垒，人工智能安全研究者。
研究智能系统在对抗环境中的安全、鲁棒与可信，
探索从模型安全走向系统与行为安全。I’m Xiaolei Liu, an AI security researcher.
I study the security, robustness and trustworthiness of intelligent systems under adversarial conditions.
刘小垒Xiaolei Liu副研究员 · 硕士生导师Associate Research Fellow
现任国家工程物理交叉科学研究中心副研究员。博士毕业于电子科技大学，师从张小松教授与朱清新教授；曾赴加州大学圣芭芭拉分校访问研究。Associate Research Fellow at the National Interdisciplinary Research Center of Engineering Physics. I received my Ph.D. from UESTC, advised by Prof. Xiaosong Zhang and Prof. Qingxin Zhu, and was a visiting scholar at UC Santa Barbara.
01 / RESEARCH
从攻击中理解风险，
在验证中建立信任。Understand the risk.
Establish the trust.
以对抗视角审视智能系统，用可复现的研究连接安全机制与真实问题。An adversarial view of intelligent systems, connecting reproducible research with real security problems.
对抗鲁棒性与隐私Robustness & Privacy
ADVERSARIAL MACHINE LEARNING
研究跨模型、跨模态的对抗迁移，查询受限条件下的黑盒攻击，以及模型反演与隐私泄露。Transferable attacks across models and modalities, query-limited black-box attacks, model inversion and privacy leakage.
相关成果Related work ↗软件与系统安全Software & Systems
DATA-DRIVEN SYSTEM SECURITY
面向软件供应链、协议与漏洞检测，探索数据驱动的安全分析方法，连接模型、软件与系统。Data-driven security analysis for software supply chains, protocols and vulnerability detection, connecting models, software and systems.
相关成果Related work ↗CURRENT PRACTICE / MAIVER
从可信模型，
走向可信行动。From trusted models
to trusted actions.
我当前也关注 AI Agent 安全与机器行为治理：当 AI 从生成内容走向调用工具、执行任务，如何理解、验证并约束它在真实系统中的行为？My current interests also extend to AI agent security and machine behavior governance: understanding, verifying and governing AI behavior as models begin using tools and acting in real systems.
了解 MAIVERExplore MAIVER02 / SELECTED WORK
以研究回应问题。Research in focus.
NEO: Navigating Entropy in Optimized Closed-Box Video Adversarial Attacks
以信息熵引导视频对抗攻击，探索查询受限条件下的模型鲁棒性。Exploring query-efficient video adversarial attacks through entropy-guided optimization.
完整作者信息、论文与补充材料见成果档案。Full author lists, papers and resources are available in the archive.浏览 29 篇论文Browse 29 publications →
03 / RECENT NOTES
近况与进展。Recent notes.
团队论文获 Inscrypt 2026 安全方向最佳学生论文奖。A team paper received the Best Student Paper Award (Security Track), Inscrypt 2026.
AWARD两篇论文被 Inscrypt 2026 录用。Two papers accepted to Inscrypt 2026.
RESEARCHPSFuzzer 被 IEEE Transactions on Software Engineering 录用。PSFuzzer accepted to IEEE Transactions on Software Engineering.
↗NEO 被 TDSC 录用；两篇论文被 ACM Multimedia 2026 录用。NEO accepted to TDSC; two papers accepted to ACM Multimedia 2026.
↗更早的进展Earlier updates
论文分别被《电子学报》与《计算机学报》录用。Papers accepted to Acta Electronica Sinica and Chinese Journal of Computers.
MADA-Attack 被 ICML 2026 录用。MADA-Attack accepted to ICML 2026.
TrajGuard 被 Findings of ACL 2026 录用。TrajGuard accepted to Findings of ACL 2026.
SEADGAT 被 IEEE Internet of Things Journal 录用。SEADGAT accepted to IEEE Internet of Things Journal.
一篇论文被《电子学报》录用。A paper accepted to Acta Electronica Sinica.
LoopLLM 与 DoBlock 被 AAAI 2026 录用。LoopLLM and DoBlock accepted to AAAI 2026.
04 / RECOGNITION
荣誉与认可。Recognition.
- 2024
- 四川省高层次人才计划Sichuan High-Level Talent Program
- 2024
- 中国电子学会青年人才托举工程计划CIE Young Elite Scientists Sponsorship Program
- 2022
- GeekPWN G-Outstanding Winner
- 2018
- 国家留学基金委公派留学奖学金China Scholarship Council Scholarship
- 2016
- 首届国家网络安全奖学金National Cybersecurity Scholarship
05 / COMMUNITY
学术共同体。Academic service.
参与专业组织与期刊服务，推动安全研究的交流与连接。Contributing to professional societies and editorial boards in security research.
专业组织Professional societies
- 中国计算机学会网络与系统安全专委会执行委员Executive Member, CCF Technical Committee on Network and System Security
- 中国电子学会网络空间安全专委会委员Member, CIE Technical Committee on Cyberspace Security
- 中国图象图形学学会数字媒体取证与安全专委会委员Member, CSIG Technical Committee on Digital Media Forensics and Security
- 中国系统工程学会网络空间安全与治理专委会委员Member, SESC Technical Committee on Cyberspace Security and Governance
- 中国计算机学会第十三届会员代表Delegate, 13th CCF Member Congress
- 四川省人工智能行业协会专家委员会专家Expert Committee, Sichuan Artificial Intelligence Industry Association
期刊编委服务Editorial service ＋
- 《通信技术》执行编委Executive Editorial Board, Communications Technology
- 《网络空间安全科学学报》青年编委Youth Editorial Board, Journal of Cyberspace Security Science
- 《信息网络安全》青年编委Youth Editorial Board, Netinfo Security
- 《重庆邮电大学学报》青年编委Youth Editorial Board, Journal of Chongqing University of Posts and Telecommunications
研究合作与人才交流Research & opportunities
欢迎围绕大模型安全、对抗鲁棒性、软件与系统安全开展合作。也欢迎有扎实基础、独立思考能力与动手热情的研究助理、实习生及研究生联系我。I welcome collaborations on LLM security, adversarial robustness, and software and system security, and inquiries from research assistants, interns and prospective graduate students.
给有意加入的同学A note to prospective students ＋
我看重扎实的算法与编程基础、清晰的表达、真诚的沟通，以及把问题想透、把工作做完整的耐心。比起只追逐论文数量，我更希望你对问题本身保持好奇，并逐步形成独立研究的能力。I value solid foundations, clear communication, intellectual honesty and the patience to see a problem through. I hope you will develop curiosity about the problem itself and the ability to pursue independent research.
我重视专注科研的时间与积极的研究氛围，鼓励使用 AI 工具提升效率，同时对方法、实验与结论负责。来信请附上简历、研究兴趣，以及一个你认真做过的项目或问题。I value focused research time and a constructive environment. AI tools are welcome, alongside responsibility for methods, experiments and conclusions. Please include your CV, research interests, and a project or question you have worked on thoughtfully.