恶意软件研究方向学生研究者，聚焦 AI Security、对抗恶意软件生成，以及长期建设更扎实的恶意软件分析与逆向工程能力。
Malware researcher focused on AI security, adversarial malware generation, and building a deeper long-term foundation in malware analysis and reverse engineering.
这个主页会逐步从“工作 / 专业 brag document”扩展成一个更完整的个人主页，同时也记录我如何学习、成长，并建立支撑长期工作的生活方式。
This profile is gradually becoming a more complete personal homepage: first as a clear work / professional brag document, and over time as a place that also reflects how I learn, grow, and build a sustainable life around the work.
我是广州大学网络空间安全学院人工智能方向硕士生，研究位于恶意软件研究与 AI 安全的交叉点。当前主要关注恶意软件生成、对抗样本生成、分布外检测，以及生成模型和强化学习在安全研究中的实际应用。
I am an MSc student at Guangzhou University, School of Cyberspace Security, working at the intersection of malware research and AI security. My current work centers on malware generation, adversarial sample generation, out-of-distribution detection, and the practical use of generative models and reinforcement learning in security research.
我重视扎实、可复用、能够持续积累的方法，希望自己逐步成长为一个既能做论文、也能做实验、还能深入恶意软件分析实践的人。
I care about doing rigorous work, building reusable research pipelines, and growing into someone who can move fluently between papers, experiments, and deeper malware analysis practice.
- 恶意软件生成：关注生成结果的可控性、真实性，以及低痕迹、高规避样本构造
Malware generation with an emphasis on controllability, realism, and low-trace, high-evasion sample construction - 对抗样本生成与安全对抗：特别关注强化学习与对抗生成的结合
Adversarial sample generation and security confrontation, especially combining reinforcement learning with adversarial generation - 逆向学习路线：作为长期方向，持续加强二进制分析与恶意软件行为理解
Reverse-engineering learning as a long-term path toward stronger binary analysis and malware behavior understanding
Unknown malware detection based on hyperspherical embedding and out-of-distribution sample detection
- 状态：在投，
IEEE IoTJ
Status: Under Review, IEEE IoTJ - 亮点：基于超球面嵌入与置信度校准的稳健 OOD 恶意软件检测
Highlight: robust OOD malware detection with hyperspherical embedding and calibrated confidence - 贡献：提出统一的超球面嵌入 + OOD 置信度框架，用于零日检测并降低误报
Contribution: a unified hyperspherical embedding + OOD confidence framework for zero-day detection with lower false positives - 方法：超球面嵌入、
SupCon、vMF建模、Mahalanobis 置信度判定
Method: hyperspherical embedding,SupCon,vMFmodeling, and Mahalanobis-confidence-based detection - 特征：融合字符串、API、opcode 等多源静态特征，并进行重点特征筛选
Features: multi-source static features across strings, APIs, and opcodes, with focused feature selection - 评估：与
KNN、Malconv、CIDER等基线进行对比
Evaluation: compared against baselines includingKNN,Malconv, andCIDER
An Evolutionary Strategy-Enhanced Conditional GAN for High-Evasion and Low-Trace Adversarial Malware Generation
- 状态：在投，期刊 / 会议待定
Status: Under Review, venue TBD - 亮点：
ES-CGAN在保持 PE 合规扰动的前提下实现高规避与低痕迹
Highlight:ES-CGANachieves high evasion with minimal, PE-compliant perturbations - 贡献：提出一个双层优化框架，在不可微字节空间中平衡规避率与隐蔽性
Contribution: a two-level optimization framework bridging non-differentiable byte space with evasion / stealth trade-offs - 方法：
CMA-ES + CGAN，面向离散 PE 扰动生成
Method:CMA-ES + CGANfor discrete PE perturbations - 结果：在 VirusTotal 上达到
79.43%峰值规避率、69.60%平均 ASR、0.9%PBR
Result: achieved79.43%peak evasion on VirusTotal,69.60%mean ASR, and0.9%PBR - 扰动设计：包含符合 PE 结构约束的 DOS / header 修改、节空间填充和新节注入
Perturbation design: PE-structure-compliant changes including DOS/header-related edits, section-space padding, and new section injection
- 状态：在投，《软件学报》
Status: Under Review, Journal of Software - 亮点：系统梳理对抗恶意软件逃逸方法与开放挑战
Highlight: a systematic review of adversarial malware evasion methods and open challenges - 贡献：构建对抗恶意软件逃逸中的威胁模型与扰动策略统一分类框架
Contribution: a unified taxonomy of threat models and perturbation strategies for adversarial malware evasion - 覆盖内容：攻击构造、扰动策略、评估指标与防御方向
Coverage: attack construction, perturbation strategies, evaluation metrics, and defense directions
- 恶意软件分析 | Malware analysis
- 安全对抗研究 | Security adversarial research
- 生成模型 | Generative models
- 强化学习 | Reinforcement learning
- 分布外检测 | OOD detection
- 数据集构建与预处理 | Dataset curation and preprocessing
- 作为对抗恶意软件研究的可复用实验支撑线持续建设
Built as a reusable experimental line supporting adversarial malware research - 已完成
CycleGAN相关实验，并建立基础训练与评估脚本
CompletedCycleGAN-related experiments and established baseline training / evaluation scripts - 持续拆解
DQN实现，强化对强化学习实验范式的理解
Continuing to unpackDQNimplementations to strengthen reinforcement learning experimentation
- 已获取奇安信
DataCon比赛数据集
Acquired the Qi An XinDataConcompetition dataset - 正在复现基线模型，并搭建更稳定的预处理与评估流程
Reproducing baseline models and building a more stable preprocessing and evaluation pipeline - 目标是形成可复用的恶意软件分析工作流
Goal: form a reusable malware analysis workflow for follow-on research
- 打磨在投论文，提升实验质量与实现细节
Refining current papers and improving experiment quality and implementation details - 复现
DataCon基线并稳定预处理 / 评估流程
ReproducingDataConbaselines and stabilizing preprocessing / evaluation workflows - 建立系统化的逆向学习路径，深化恶意软件行为分析能力
Building a systematic reverse-engineering learning path for deeper malware behavior analysis
- 在工作之外继续认真学习，慢慢扩展自己的视野
Learning beyond work, slowly and deliberately - 建立支持长期深度工作的生活技能与可持续习惯
Building life skills and sustainable routines that support deep work - 发展研究之外的创造性与反思性实践
Developing creative and reflective practices outside research - 记录细小但真实的个人成长
Documenting small but meaningful personal growth over time
这些页面目前还是占位，后续会逐步从研究型简介扩展成更完整的个人主页。
These are placeholders for future pages as this profile evolves from a research profile into a broader personal homepage.
-
广州大学，网络空间安全学院
Guangzhou University, School of Cyberspace Security人工智能硕士，
2025-2027（预计）
MSc in Artificial Intelligence,2025-2027(expected) -
海南大学，土木建筑工程学院
Hainan University, College of Civil Engineering & Architecture工程管理本科，
2017-2021
BEng in Engineering Management,2017-2021