EM Side-Channel Based Hardware Trojan Detection Using Deep SVDD for Unsupervised Anomaly Identification

Abstract

The globalization of integrated-circuit design and fabrication has increased the risk of malicious hardware modifications. This paper presents a non-invasive electromagnetic side-channel framework for hardware Trojan detection using Deep Support Vector Data Description (Deep SVDD). Electromagnetic emissions from a device under test are captured with a near-field probe, conditioned, digitized, and transformed into the frequency domain using a Hann window and Fast Fourier Transform. The magnitude-squared spectrum is computed, and dominant peaks near the operating clock and its harmonics are selected to form a compact feature vector. Only Trojan-free samples are used during offline training. Deep SVDD learns a compact latent representation of normal behavior by minimizing the distance of mapped feature vectors from a hypersphere center. During online operation, the anomaly score of a new sample is compared with a learned threshold to classify the device as normal or suspicious. MATLAB is used for dataset preparation and model training, while the online signal-processing and scoring pipeline is structured for Verilog implementation. The proposed approach does not require labeled Trojan examples, supports the detection of previously unseen modifications, and offers a practical path toward real-time trust verification of integrated circuits