Abstract
Hardware Trojan detection is critical for ensuring integrated circuit security, particularly against instruction-activated hardware Trojans—which possess formidable stealth capabilities and targeted attack potential. Existing machine learning-based detection methods suffer from multiple limitations: insufficient feature separation, difficulty distinguishing normal noise from faint Trojan activation signals, and limited detection capability against instruction-activated Trojans. To address these challenges, this paper proposes the scale time attention detection (STAD) model, a semi-supervised anomaly detection method designed to resolve the issue of undefined anomaly boundaries in unsupervised learning. The model innovatively integrates multi-scale feature extraction with temporal attention mechanisms. Its core architecture comprises three parallel multi-scale temporal convolutional network (TCN) branches, capturing local temporal dependencies through sequences with varying dilation rates. The temporal attention mechanism applies local convolutional operations to learn adaptive importance weights for different time steps, leveraging the multi-scale temporal features encoded by the TCN branches. The dual-pooling layer preserves richer feature information, while the fully connected layer is used to compute anomaly scores. Experiments on the SAKURA-G development board demonstrate that the STAD detection model achieves favorable classification results using data from AES and RSA series hardware Trojans embedded in the Trust-Hub benchmark.