DED: A Dynamic-Enhancing Defense Framework Against Model Extraction Attacks in Industrial Internet of Things
发布时间：2026-10-18 点击次数：
姚羽（教授）
+
- 博士生导师 硕士生导师
- 所在单位：计算机科学与工程学院
- 职务：复杂网络系统安全保障技术教育部工程研究中心主任
- 学历：博士研究生毕业
- 性别：男
- 联系方式：
- 学位：博士
- 毕业院校：东北大学
- 所属院系：计算机科学与工程学院
- 学科：
计算机应用技术
计算机软件与理论
计算机系统结构
- 论文名称：DED: A Dynamic-Enhancing Defense Framework Against Model Extraction Attacks in Industrial Internet of Things
- 发表刊物：IEEE Transactions on Industrial Informatics
- 影响因子：9.8
- 摘要：In the industrial Internet of Things (IIoT) environment, intelligent models are typically trained on proprietary industrial data, which endows them with substantial intellectual property and commercial value. To safeguard these assets, model owners usually restrict access to closed box interfaces. However, model extraction attacks allow adversaries to steal models without internal access, posing serious threats to IIoT security. Existing defense methods are mainly designed for general domains and fail to consider the unique characteristics of industrial applications. When directly applied to IIoT, they often result in degraded performance or limited applicability. Specifically, previous perturbation-based strategies can partially mitigate attacks but inevitably harm model performance, which is unacceptable in reliability-critical IIoT scenarios; meanwhile, retraining-based defenses are impractical since IIoT models are usually impossible to retrain. To address these challenges, this letter proposes a bypass-style defense method against model extraction attacks tailored for IIoT applications. The proposed approach leverages a deep hash model and a hash-conditional diffusion model to extract subtle features from industrial samples for efficient filtering, and achieves dynamic defense by constructing a malicious sample repository. Experimental results on three representative IIoT datasets demonstrate that our method outperforms existing approaches in terms of defense effectiveness while preserving model performance.
- 关键字：In the industrial Internet of Things (IIoT) environment, intelligent models are typically trained on proprietary industrial data, which endows them with substantial intellectual property and commercial value. To safeguard these assets, model owners usually restrict access to closed box interfaces. However, model extraction attacks allow adversaries to steal models without internal access, posing serious threats to IIoT security. Existing defense methods are mainly designed for general domains and fail to consider the unique characteristics of industrial applications. When directly applied to IIoT, they often result in degraded performance or limited applicability. Specifically, previous perturbation-based strategies can partially mitigate attacks but inevitably harm model performance, which is unacceptable in reliability-critical IIoT scenarios; meanwhile, retraining-based defenses are impractical since IIoT models are usually impossible to retrain. To address these challenges, this letter proposes a bypass-style defense method against model extraction attacks tailored for IIoT applications. The proposed approach leverages a deep hash model and a hash-conditional diffusion model to extract subtle features from industrial samples for efficient filtering, and achieves dynamic defense by constructing a malicious sample repository. Experimental results on three representative IIoT datasets demonstrate that our method outperforms existing approaches in terms of defense effectiveness while preserving model performance.
- 论文类型：SCI JCR Q1
- 备注：https://ieeexplore.ieee.org/abstract/document/11737564
- 学科门类：工学
- 文献类型：JCR 一区
- 一级学科：计算机科学与技术
- 是否译文：否