Hybrid GAN-RL Architecture for Hardware Trojan Detection in Gate-Level Netlists

Douar M. O. I., Esirci F. N., KAKIŞIM A.

8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2026, Ankara, Türkiye, 21 - 23 Mayıs 2026, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/ichora69329.2026.11537257
  • Basıldığı Şehir: Ankara
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: Deep Q-Network, Explainable AI, Gate-level, Generative Adversarial Networks, Hardware Trojan, Imbalanced Data, Pre-silicon, Reinforcement Learning, Trojan Detection
  • Yıldız Teknik Üniversitesi Adresli: Evet

Özet

The modern integrated circuit (IC) supply chain involves multiple stages, such as design, fabrication, and testing, which are often distributed across different companies, countries, and regions. This distributed structure increases the risk of Hardware Trojans (HTs), which can be maliciously inserted at different stages of the IC supply chain, making pre-silicon detection critical before compromised designs progress further. Although learning based methods have shown potential for pre-silicon HT detection, their performance can be limited by class imbalance and false positive rates (FPR). In particular, high FPR may lead to clean circuit designs being incorrectly flagged as suspicious, reducing the practicality of such methods. To address these challenges, this paper proposes a Hybrid GAN-RL architecture for pre-silicon hardware Trojan detection from gate-level netlists. The architecture utilizes a conditional tabular generative adversarial network (CTGAN) to mitigate class imbalance by generating synthetic circuit samples for the underrepresented class (clean circuits). It then performs Trojan detection using a DQN-based reinforcement learning (RL) agent designed to reduce FPR. Experimental results demonstrate improved performance over baseline methods, achieving up to 99.22% accuracy with a false positive rate of $2.22\%$ under cross-validation. These results suggest that combining generative data augmentation with reinforcement learning can improve the reliability of pre-silicon Hardware Trojan detection under imbalanced data conditions.