[Paper Review] Node-wise Hardware Trojan Detection Based on Graph Learning

Kento Hasegawa, K. Yamashita+4 more|arXiv (Cornell University)|Dec 4, 2021

Physical Unclonable Functions (PUFs) and Hardware Security4 citations

TL;DR

This paper proposes NHTD-GL, a node-wise hardware Trojan (HT) detection method using graph learning (GL) that automatically extracts HT features from gate-level netlists without heuristic feature engineering. It achieves 0.998 detection accuracy and outperforms state-of-the-art methods by effectively capturing HT-specific features through graph neural networks (GNNs), validated on both benchmark and randomly generated HTs.

ABSTRACT

In the fourth industrial revolution, securing the protection of the supply chain has become an ever-growing concern. One such cyber threat is a hardware Trojan (HT), a malicious modification to an IC. HTs are often identified in the hardware manufacturing process, but should be removed earlier, when the design is being specified. Machine learning-based HT detection in gate-level netlists is an efficient approach to identify HTs at the early stage. However, feature-based modeling has limitations in discovering an appropriate set of HT features. We thus propose NHTD-GL in this paper, a novel node-wise HT detection method based on graph learning (GL). Given the formal analysis of HT features obtained from domain knowledge, NHTD-GL bridges the gap between graph representation learning and feature-based HT detection. The experimental results demonstrate that NHTD-GL achieves 0.998 detection accuracy and outperforms state-of-the-art node-wise HT detection methods. NHTD-GL extracts HT features without heuristic feature engineering.

Motivation & Objective

• Address the limitations of heuristic feature engineering in existing hardware Trojan (HT) detection methods, which struggle to adapt to new or unknown HT variants.
• Enable practical, early-stage HT detection during the IC design phase by focusing on gate-level netlists, where HTs are more likely to be inserted via third-party IPs or EDA tools.
• Bridge the gap between domain knowledge of HT structural features and the representation learning capability of graph neural networks (GNNs) to improve detection robustness.
• Demonstrate that GNN-based models can generalize to HTs with unknown or unseen features, reducing reliance on pre-defined feature sets.
• Provide a theoretically grounded, scalable, and high-accuracy solution for node-wise HT detection in real-world IC supply chains.

Proposed method

• Represent gate-level netlists as heterogeneous graphs, where nodes correspond to circuit elements (e.g., logic gates) and edges represent interconnections (wires).
• Apply Graph Attention Networks (GAT), Message Passing Neural Networks (MPNN), and Graph Isomorphism Networks (GIN) as the core GNN architectures to learn node-level representations.
• Leverage domain knowledge of HT characteristics—such as trigger and payload structures—to guide the design of the graph learning framework and interpret model behavior.
• Train the GNN models end-to-end on labeled netlist data (normal vs. Trojan nodes) to predict node-wise HT presence without requiring manual feature extraction.
• Use a multi-head attention mechanism in GAT to dynamically weigh the importance of neighboring nodes during message passing, enhancing feature discrimination for HT nodes.
• Evaluate model performance using standard metrics (recall, precision, F1-score, accuracy) on both Trust-HUB benchmark datasets and randomly generated HT-infested circuits.

Experimental results

Research questions

• RQ1 Can graph learning effectively detect hardware Trojans at the node level in gate-level netlists without relying on manually engineered features?
• RQ2 How well do GNN-based models generalize to hardware Trojans with unknown or unseen structural features not present in training benchmarks?
• RQ3 What specific features do GNNs learn that enable high-accuracy HT detection, and how do they relate to known HT characteristics from domain knowledge?
• RQ4 Does the proposed method outperform existing state-of-the-art node-wise HT detection techniques in terms of accuracy and robustness across diverse HT types?
• RQ5 Can the model maintain high performance when detecting small, stealthy Trojans that constitute less than 5% of total nodes in a circuit?

Key findings

• NHTD-GL achieves a detection accuracy of 0.998 on the Trust-HUB benchmark dataset, significantly outperforming existing node-wise HT detection methods.
• The method achieves a recall of 0.890 and precision of 0.978, with an F1-score of 0.921, demonstrating strong balance between sensitivity and specificity.
• On randomly generated HTs with unknown features, NHTD-GL maintains high performance with a recall of 0.773, precision of 0.843, and accuracy of 0.995, proving generalization capability.
• GNN-based models (e.g., GAT, MPNN, GIN) consistently outperform the traditional random forest method from [30], which relies on handcrafted features and fails on unseen HT types.
• The GNN models successfully extract HT-relevant features automatically, confirming that graph learning can overcome the limitations of heuristic feature engineering.
• Theoretical analysis confirms that GNNs capture structural features aligned with known HT patterns—such as trigger and payload configurations—validating the model’s interpretability and reliability.