Efficient hardware Trojans detection through multi-scale features fusion of thermal radiation maps

Abstract

Thermal radiation (TR) has emerged as a promising technique for post-silicon hardware Trojan (HT) detection. However, existing TR-based methods rely on complex analyses of thermal radiation maps (TRMs), resulting in significant time overhead. To this end, we propose the TURBO framework, which performs TRMs feature fusion via the U-Net architecture for batch detection of high-volume ICs. TURBO leverages U-Net’s ability to extract and integrate multi-scale TR features, effectively distinguishing HTs from vacant regions. Experiments show that TURBO achieves over 95% detection rate for pixel-level HTs under 130 nm and 28 nm nodes, with a runtime of 4 minutes for 10⁵ pixels.