晨光
暗夜
晨光
极光
Bilingual Paper Reading · 中英对照精读

基于图神经网络的硬件木马检测:免「黄金参考」覆盖 RTL 与门级网表

准大一 · 集成电路设计与集成系统 × 硬件安全 × AI —— 硬件木马检测精读材料
原文:arXiv:2204.11431 2022年4月25日发布 arXiv 预印本(cs.CR / cs.AR) 硬件木马 × 图神经网络 × 供应链安全 附英文摘要朗读音频

一、论文档案

英文标题Hardware Trojan Detection using Graph Neural Networks
中文标题基于图神经网络(GNN)硬件木马检测
作者罗津·亚萨伊, 卢克·陈, 于世远, 穆罕默德·阿卜杜拉·阿尔·法鲁克(美国加州大学尔湾分校,加州大学尔湾分校)
发布时间2022年4月25日(v1)|分类:cs.CR(密码学与安全)、cs.AR(硬件架构)
一句话概括不需要「黄金参考电路」,把 RTL 和门级网表都画成数据流图(DFG),用图神经网络学电路行为,97% 召回率、21.1ms 就能揪出没见过的硬件木马。
💡 为什么选这篇给你:① 硬件木马是芯片供应链安全的核心威胁,直接关系国防、医疗、关键基础设施的芯片可信度;② 思路新——用 GNN 把「图结构」当电路行为的载体,彻底摆脱对「黄金参考」的依赖;③ 结果数字漂亮(97% 召回率、毫秒级检测),故事完整、方法可复现,非常适合入门 AI+硬件安全。

二、核心术语表(先扫一遍再读正文)

英文术语中文大白话解释
Hardware Trojan (HT)硬件木马被恶意植入芯片的「后门电路」,平时休眠、特定条件触发后泄密/改功能/搞破坏。
golden reference circuit黄金参考电路公认「绝对干净」的原始电路,传统检测方法拿它当对照标准;本文的方法不需要它。
3PIP第三方知识产权核向外部供应商购买的 IP 核(如 Verilog 模块),可复用省钱,但可能被植入木马。
RTL寄存器传输级用 Verilog/VHDL 描述的数字电路设计层级,描述「寄存器之间怎么传数据、做什么运算」。
gate-level netlist门级网表综合之后由与门、或门、触发器组成的电路连线清单,比 RTL 更接近物理实现。
Data Flow Graph (DFG)数据流图把电路画成图:节点是运算/元件,边是数据流动方向——图神经网络的输入。
Graph Neural Network (GNN)图神经网络专门处理「图」数据的神经网络,能学习节点之间如何互相影响。
pre-silicon流片前芯片还没制造之前的验证阶段,此时发现木马改起来最便宜。
payload / trigger木马载荷 / 触发器trigger 是「开关」:罕见条件满足才激活;payload 是「炸弹」:被激活后干的坏事。
recall / F1-score召回率 / F1 分数recall=真木马里被找出来的比例;F1 是精确率与召回率的调和平均,越高越均衡。
TrustHub benchmarkTrustHub 基准学术界公认的硬件木马测试集(如 AES-T600/T700),用来公平比较检测方法。
SoC系统级芯片把 CPU、内存、外设控制器等集成到一颗芯片上的片上系统。
3P-EDA tools第三方 EDA 工具第三方厂商提供的设计自动化软件,如果不可信,也可能在综合时偷偷塞入木马。

三、摘要中英对照(精读核心)

🎧 音频在文末,可先听一遍原文再读;每个英文句都配了逐句翻译。

摘要 Abstract

EN · 原文
The globalization of the Integrated Circuit (IC) supply chain has moved most of the design, fabrication, and testing process from a single trusted entity to various untrusted third-party entities around the world.
CN · 翻译
集成电路(IC)供应链的全球化,把设计、制造、测试的大部分环节从「单一可信实体」转移到了世界各地不可信的第三方实体手中。
EN · 原文
The risk of using untrusted third-Party Intellectual Property (3PIP) is the possibility for adversaries to insert malicious modifications known as Hardware Trojans (HTs).
CN · 翻译
使用不可信第三方知识产权核(3PIP)的风险在于:攻击者可能植入恶意修改,即硬件木马(HT)
EN · 原文
These HTs can compromise the integrity, deteriorate the performance, and deny the functionality of the intended design.
CN · 翻译
这些木马可以破坏设计完整性、降低性能、甚至让功能失效
EN · 原文
Various HT detection methods have been proposed in the literature; however, many fall short due to their reliance on a golden reference circuit, a limited detection scope, the need for manual code review, or the inability to scale with large modern designs.
CN · 翻译
文献中已有多种木马检测方法,但很多都掉链子:要么依赖黄金参考电路、要么检测范围有限、要么需要人工审代码、要么无法扩展到现代大规模设计。
EN · 原文
We propose a novel golden reference-free HT detection method for both Register Transfer Level (RTL) and gate-level netlists by leveraging Graph Neural Networks (GNNs) to learn the behavior of the circuit through a Data Flow Graph (DFG) representation of the hardware design.
CN · 翻译
我们提出一种免黄金参考的木马检测新方法,同时覆盖寄存器传输级(RTL)门级网表:用图神经网络(GNN),通过硬件设计的数据流图(DFG)表示来学习电路行为。
EN · 原文
We evaluate our model on a custom dataset by expanding the Trusthub HT benchmarks.
CN · 翻译
我们在扩展 TrustHub 木马基准构建的自建数据集上评估模型。
EN · 原文
The results demonstrate that our approach detects unknown HTs with 97% recall (true positive rate) very fast in 21.1ms for RTL and 84% recall in 13.42s for Gate-Level Netlist.
CN · 翻译
结果表明:对未知木马,RTL 上以 97% 召回率21.1ms 极速检出;门级网表上以 84% 召回率13.42s 检出。

关键词 Keywords:Hardware Trojan Detection 硬件木马检测 | Graph Neural Networks 图神经网络 | Data Flow Graph 数据流图 | Golden Reference-Free 免黄金参考 | RTL / Gate-Level Netlist RTL / 门级网表

四、引言精选(为什么这个问题重要)

① 背景:SoC 设计外包与「信任第三方」的代价

EN · 原文
The scale and complexity of modern System-on-Chip (SoC) designs have made it increasingly challenging and expensive for chip manufacturers to design, fabricate, and test every component in-house. The time to market pressure and resource constraints have pushed SoC designers to outsource hardware designs and use Third-Party Electronic Design Automation (3P-EDA) tools and Intellectual Property (IP) cores from various vendors worldwide. Using Third-Party IPs (3PIP) can be cost-effective due to the re-usability of IP cores so that chip manufacturers can reallocate their resources to meet market demands. However, the security and trustworthiness of 3PIPs are not always guaranteed, and reliance on untrusted IPs and EDA tools greatly raises the risks of HT insertion by rogue entities in the IC supply chain.
CN · 翻译
现代片上系统(SoC)设计规模大、复杂度高,芯片厂商自己设计、制造、测试所有部件越来越难、越来越贵。上市时间和资源压力逼着设计师外包硬件设计、使用全球供应商的第三方 EDA 工具和 IP 核。用 3PIP 省钱(IP 可复用),但安全与可信无法保证——依赖不可信的 IP 和 EDA 工具,大大增加了供应链中恶意实体植入木马的风险。

② 威胁:一个未被发现的木马可能是「致命的」

EN · 原文
HT refers to an intentional and malicious modification of an IC that is usually designed to leak the information, change the functionality, degrade the performance, or deny the service of the chip. Due to the wide applications of ICs in military systems, critical infrastructures, medical devices, etc., the consequences of an undetected HT in a chip can be life-threatening. For example, an actual demonstration of the HT threat occurred in 2007, when a suspected nuclear installation in Syria was bombed by Israeli jets because Syrian radar was disabled by a remote kill switch backdoor in its commercial off-the-shelf microprocessor [2]. In 2012, an undocumented hardware backdoor was found in the Actel/Microsemi ProASIC3 chips used in military-grade FPGAs [3] that allowed the extraction of secret keys, enabling an adversary to modify the chip's configurations and gain full control of the chip. Furthermore, it is projected that the global semiconductor IP market will reach 7.3 Billion by 2025, with a compounded annual growth rate of 5.5% from 2020-2025 [4], and the security concerns about untrusted IPs can significantly damage the market.
CN · 翻译
硬件木马是对 IC 的蓄意恶意修改,通常用于泄密、改功能、降性能或拒绝服务。IC 广泛用于军事系统、关键基础设施、医疗设备等,漏检一个木马的后果可能是致命的。真实案例:2007 年叙利亚疑似核设施被以色列战机轰炸,原因是其商用微处理器里的「远程击杀开关」后门瘫痪了叙方雷达;2012 年,军用级 FPGA(Actel/Microsemi ProASIC3)中被发现未记录的硬件后门,可窃取密钥、让攻击者完全控制芯片。此外,预测全球半导体 IP 市场到 2025 年将达 73 亿美元(2020-2025 年复合年增长率 5.5%),对不可信 IP 的安全担忧会严重损害这个市场。

③ 攻防拉锯:加州大学尔湾分校 → SMC → FANCI → DeTrust → Mero

EN · 原文
The HT detection problem has always been a back-and-forth tug-of-war. HTs are stealthy by design and are composed of a payload and trigger. They are usually very small and inactive with minimal effects on the chip until the trigger circuit is activated under very rare circumstances and triggers the payload to perform its malicious activities. One of the earliest defense mechanisms, [6] proposed a novel Unused Circuit Identification (加州大学尔湾分校) technique that identifies suspicious circuitry not being used or activated during design verification. However, the authors in [7] later designed a new type of HT called Stealthy Malicious Circuits (SMC), which could bypass 加州大学尔湾分校 by hiding HT in nearly-unused logic. Further, FANCI [8] was successful in detecting SMC by identifying the low control value exhibited by the nearly-unused logic, but it was later defeated by DeTrust [9], which designed a new class of HTs with stealthy implicit triggers. Mero[10] generates test vectors that are capable of activating HTs with low trigger probabilities for detection. Still, it fails to generate test vectors that can activate "hard-to-trigger" HTs with trigger probabilities less than 10-6 [11].
CN · 翻译
木马检测是一场永不停歇的拉锯战。木马天生隐蔽,由 payload(载荷)和 trigger(触发器)组成,通常极小且休眠,直到触发器在极罕见条件下被激活。最早的防御之一 加州大学尔湾分校 技术能识别「验证中从未被使用/激活」的可疑电路;但随后 SMC 木马把木马藏在近乎未用的逻辑里绕过 加州大学尔湾分校;FANCI 通过识别近未用逻辑的低控制值成功检出 SMC,又被 DeTrust 的「隐式触发器」木马击败;Mero 能生成激活低触发概率木马的测试向量,却对付不了触发概率低于 10⁻⁶ 的「难触发」木马。典型的一攻一防、道高一尺魔高一丈。

④ 现状缺口:基于图的方法仍局限于「已知木马」

EN · 原文
Among the more recent works that use Trusthub benchmarks, [12] is supposed to detect HTs that leak sensitive data such as secret keys in cryptographic cores as in the AES-T600 benchmark. However, it fails to detect another similar HT benchmark, AES-T700. Later, [13] identifies different data leaking HTs (e.g., AES-T600 and AES-T700) by adding data leaking as an additional security property for model checking. However, it fails for HTs that execute other malicious functions such as chip degradation (e.g., the AES-500 benchmark) rather than data leakage. A recent paper [14] showed that modeling the hardware design as a graph can be beneficial in the hardware security domain. However, the HT detection scope of [15, 14] based on graph similarity algorithms are limited to known HTs in the method's library.
CN · 翻译
近年基于 TrustHub 基准的工作:某方法能检出 AES-T600(泄密型木马),却漏掉相似的 AES-T700;后续方法把「数据泄露」加为模型检验的安全属性,能同时检出 T600/T700,但又对付不了 AES-500 这类「芯片退化」型木马。最近有论文证明把硬件设计建模成图很有用,但基于图相似度算法的方法,检测范围仅限于方法库里已知的木马——这正是本文要突破的缺口。
💡 这是全文最有味道的一句"The HT detection problem has always been a back-and-forth tug-of-war."——硬件安全领域的第一性认知:检测与反检测互相进化,所以方法必须「免黄金参考 + 能泛化到未知木马」,而不是记住已知木马的样子。

五、论文贡献(3 个要点)

EN · 原文
1. Golden reference-free detection at two levels. We propose a novel golden reference-free HT detection method for both Register Transfer Level (RTL) and gate-level netlists by leveraging Graph Neural Networks (GNNs) to learn the behavior of the circuit through a Data Flow Graph (DFG) representation of the hardware design.
CN · 翻译
1. 免黄金参考的双层次检测。提出免黄金参考的新方法,同时覆盖 RTL 与门级网表:用 GNN 通过数据流图表示学习电路行为。
EN · 原文
2. Automatic feature extraction. Our method automatically extracts the features of graphs and learns the behavior of the hardware design.
CN · 翻译
2. 自动特征提取。方法自动提取图的特征并学习硬件设计的行为——不需要专家手工设计特征。
EN · 原文
3. Fast and accurate detection of unknown HTs. The results demonstrate that our approach detects unknown HTs with 97% recall (true positive rate) very fast in 21.1ms for RTL and 84% recall in 13.42s for Gate-Level Netlist.
CN · 翻译
3. 快速准确的未知木马检测。结果证明:RTL 上 97% 召回率、21.1ms;门级网表上 84% 召回率、13.42s——对从未见过的木马依然有效。

六、结论中英对照

EN · 原文
This paper proposes a novel golden reference-free approach to find unknown HT in both RTL codes and gate-level netlists. We generate DFGs for both RTL and netlist codes and employ the GNN to construct two models that infer the presence of HT from the generated graphs in RTL and gate-level netlist. Our method automatically extracts the features of graphs and learns the behavior of the hardware design. Our model is trained and tested on a DFG dataset created by expanding the Trustub benchmarks. The RTL results indicate that the proposed method discovers HT with 97% recall and 94% F1-score very fast in 21.1ms. The gate-level netlist results indicate an 84% recall and 86% F1-score with an average detection time of 13.72 seconds.
CN · 翻译
本文提出一种免黄金参考的新方法,用于在 RTL 代码和门级网表中发现未知木马:为两者生成 DFG,并用 GNN 构建两个模型,从图里推断木马是否存在。方法自动提取图特征、学习硬件设计行为。模型在扩展 TrustHub 基准得到的 DFG 数据集上训练和测试。RTL 结果:97% 召回率、94% F1 分数、21.1ms;门级网表结果:84% 召回率、86% F1 分数、平均检测时间 13.72 秒
EN · 原文
This research was supported by the Office of Naval Research (ONR), award number N00014-17-1-2499. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of our funding agencies.
CN · 翻译
本研究由美国海军研究办公室(ONR)资助(项目号 N00014-17-1-2499)。本文观点、发现、结论或建议均属作者,不代表资助机构立场。

七、编者解读:这篇论文到底讲了什么(大白话版)

  1. 问题:芯片设计越来越依赖外购的 IP 核和第三方工具,攻击者可能在里面埋「硬件木马」——平时睡觉、关键时刻泄密或搞破坏。传统检测方法要么需要「黄金参考电路」(公认干净的版本,现实中往往没有),要么只能识别已知木马、要么靠人肉审代码,都撑不住现代大芯片。
  2. 做法:把电路(RTL 代码或门级网表)自动转成数据流图(DFG)——节点是运算、边是数据流向——然后用图神经网络(GNN)在这张图上学习「干净电路长什么样、木马电路长什么样」,直接对图分类。整个过程不需要黄金参考,也不需要人设计特征。
  3. 结果:RTL 上 97% 召回率、21.1ms 出结果;门级网表上 84% 召回率、平均 13.72 秒。F1 分数分别是 94% 和 86%。注意:摘要里门级网表写的是 13.42s,结论里写的是 13.72 秒——原文两处数字不一致,我们照实保留。
  4. 最值钱的观点:检测木马不能「记住已知木马的样子」,而要学「电路行为的一般规律」。把设计表示成图 + 让 GNN 自动学特征,才能对付不断进化的新木马。
  5. 工程意义:流片前(pre-silicon)发现木马成本最低。对芯片设计公司来说,这种自动化、免黄金参考的检测可以嵌入 EDA 流程,在综合、验证阶段就拦下恶意逻辑,守护军用、医疗、基础设施芯片的可信度。
🎯 对保研的启示:这篇论文示范了「跨领域借力」——把 AI 圈成熟的图神经网络,迁移到硬件安全这个传统领域。复试时若能讲清「为什么传统方法不行 → 图表示解决了什么 → 结果数字怎么读」,就是一个完整且有迁移价值的科研叙事。

八、给准大一的阅读路线图 & 延伸方向

📖 怎么读这篇论文(三遍法)

  1. 第一遍(10 分钟):只读摘要和术语表,回答三个问题——问题是什么?方法是什么?结果是什么?(97% / 21.1ms 这两个数字要记住)
  2. 第二遍(20 分钟):读引言 + 结论,重点体会「为什么必须免黄金参考」以及攻防拉锯的历史(加州大学尔湾分校→SMC→FANCI→DeTrust)。
  3. 第三遍(30 分钟):把引言里 4 段中英对照反复读,跳过所有引用编号 [2][3]…,只读文字;遇到不懂的术语回查术语表。

🚀 这个方向你能延伸做什么

九、英文摘要朗读(练听力用)

先盲听一遍→再看对照稿→再听一遍。目标是听出每个数字(97% recall、21.1ms、13.42s)和术语(golden reference、Data Flow Graph、Graph Neural Networks)。