Uncertainty-Aware Hardware Trojan Detection Using Multimodal Deep Learning

TL;DR

The paper tackles hardware Trojan detection in a zero-trust fabless pipeline where labeled Trojan-infected data are scarce. It introduces NOODLE, a multimodal deep-learning framework that fuses graph-based RTL representations and Euclidean/tabular circuit data, augmented by GANs to address data scarcity. Uncertainty quantification is achieved via conformal prediction and p-value fusion, enabling calibrated and risk-aware decisions; early and late fusion strategies are evaluated. Experiments on TrustHub RTL-level data show improved probabilistic accuracy and strong discriminative performance, with a ROC-AUC of 0.928 and late-fusion Brier-score of 0.1589, illustrating robustness to small datasets and the practicality of uncertainty-aware multimodal HT detection.

Abstract

The risk of hardware Trojans being inserted at various stages of chip production has increased in a zero-trust fabless era. To counter this, various machine learning solutions have been developed for the detection of hardware Trojans. While most of the focus has been on either a statistical or deep learning approach, the limited number of Trojan-infected benchmarks affects the detection accuracy and restricts the possibility of detecting zero-day Trojans. To close the gap, we first employ generative adversarial networks to amplify our data in two alternative representation modalities, a graph and a tabular, ensuring that the dataset is distributed in a representative manner. Further, we propose a multimodal deep learning approach to detect hardware Trojans and evaluate the results from both early fusion and late fusion strategies. We also estimate the uncertainty quantification metrics of each prediction for risk-aware decision-making. The outcomes not only confirms the efficacy of our proposed hardware Trojan detection method but also opens a new door for future studies employing multimodality and uncertainty quantification to address other hardware security challenges.

Figures (5)

• Figure 1: NOODLE framework: The input consists of an RTL file (Verilog), which undergoes conversion into both graph and Euclidean representations, and then input into a multimodal deep learning classifier. This classifier yields a decision indicating whether the circuit is Trojan-infected or Trojan-free.
• Figure 2: NOODLE's Brier score (a) Early fusion (b) Late fusion
• Figure 3: NOODLE's confidence calibration curve
• Figure 4: NOODLE's ROC-AUC curve under late fusion
• Figure 5: NOODLE's radar plot for consolidated metrics