基于低秩矩阵分解和大间隔学习的硬件木马检测算法研究

The security of integrated circuits (IC) play a very significant role on China's military, economy, communication and other industries. At present, the key IC designed at home has to manufacture on the controlled production line, which leads to the threat of hardware Trojan. The detection theory and methods have become a hot research topic in the field of integrated circuit. Focusing on the residual error and inefficiency of traditional artificial method, machine learning methods for large scaled integrated circuit are proposed considering the process variations, the environmental noise and the measurement noise. The destructive techniques and non-destructive technologies of traditional detection methods are studied respectively. In view of the map images achieved by reverse engineering, we analyze the noise and describe it based on the mixed Gaussian distribution, and a model based on low rank matrix decomposition is constructed. For side channel signal achieved by non-destructive technology, we consider the process variety, predict them with low cost, and propose an algorithm with high discrimination and max margin.

集成电路的安全隐患使得我国军事、经济、通信等各个行业难以得到切实有力的安全保障。目前我国自主设计的关键集成电路普遍在非受控工艺线加工制造，这一现状导致了集成电路存在硬件木马的威胁，因此硬件木马检测技术的研究具有使命意义。针对传统人工检测技术不可避免的漏检误差和效率低下的问题，本项目利用机器的精准计算和不知疲惫的特性，考虑工艺波动、环境噪声、测量噪声的影响，研究大规模集成电路硬件木马的智能检测算法。分别针对传统检测方法中的破坏性技术与非破坏性技术展开研究。针对破坏性技术中的批量版图图像，分析实际噪声组成，提出基于混合高斯分布的噪声描述方法，建立基于低秩矩阵分解的模型。针对非破坏性技术中的旁路信号这种微弱信号，充分考虑工艺波动的影响，建立低成本的工艺波动预测方法，消除旁路信号中工艺波动的影响，建立具有判别性的大间隔算法。

随着集成电路从设计到应用的全球化发展，集成电路面临着硬件木马的安全威胁。硬件木马检测准确率受硬件木马检测分辨率和信噪比影响很大。以报道的检测方法往往针对仿真电路，没有实际电路的检测结果。针对我国自主设计的集成电路在非受控工艺线上加工面临的安全隐患，以及现有硬件木马检测方法受噪声影响检测准确率不高的问题，项目开展了基于机器学习的集成电路硬件木马检测算法研究。针对实际电路的版图图像，设计了低秩矩阵分解的硬件木马检测方法，检测准确率达到100%。针对实际电路的旁路信号，设计了大间隔学习的硬件木马检测方法，可以检测出硬件木马分辨率只有1E-5的硬件木马。项目形成了三组硬件木马检测数据，包括一组仿真数据，两组实测数据。研究成果可以用在集成电路的安全性分析，为保障我国集成电路安全性提供了一定的理论支撑，具有良好的社会效益。