TAMIS: Golden-Model-Free EM Trojan Detection via Temperature-Aware Multi-Instance Segmentation and Clustering

We present a Golden-Model-free framework for hardware Trojan (HT) detection that treats temperature not as noise to remove but as physical context that organizes electromagnetic (EM) variability. At the core is Temperature-Aware Multi-Instance Segmentation (TAMIS), which synchronously acquires EM and temperature, forms short overlapping sub-windows using a hot-start sampling strategy to focus on high-activity transitions, applies changepoint-weighted spectral statistics to preserve brief activations, and augments each instance with compact thermo-temporal descriptors. Unsupervised clustering on these instances, followed by instance-to-trace aggregation, yields trace-level decisions without labels or trusted devices. On 26 Trust-Hub AES benchmarks, the pipeline attains an average F1-score of 97.02% in a fully Golden-Model-free setting. Ablation studies confirm that the hot-start strategy contributes a ∼13 pp gain in F1 in a controlled ablation on AES128-T100 by capturing transient thermal events. While mean aggregation performs well on average, it fails to detect sparse triggers, indicating a conditional advantage. By unifying EM and temperature in a single, interpretable preprocessing pipeline, TAMIS advances multi-parameter HT screening toward scalable, environment-aware, and Golden-Model-free operation. All results are obtained fromphysical measurementson a Digilent Arty A7 FPGA platform using Trust-Hub AES benchmarks, with synchronized EM and on-chip (XADC) temperature acquisition.