Hybrid Deep Neural Architectures for Hardware Trojan Identification via Side-Channel Power Analysis

Integrated circuits (ICs) are vulnerable to serious security risks from hardware Trojans (HTs), which can cause malfunctions, data leaks, or system failure. Conventional detection techniques like side-channel analysis and functional testing frequently need for destructive access or a golden reference model. In this work, a non-invasive deep learning framework that uses power side-channel signals instead of a golden chip is proposed. Long Short-Term Memory (LSTM), Autoencoder integrated with LSTM (AE_LSTM), and Convolutional Neural Networks (CNN) are the three deep learning models that are used and assessed. After preprocessing power traces into 1D arrays and normalizing them, the models are trained to identify if an IC is "TrojanEnabled" or "TrojanDisabled." Classification accuracies of 89% (LSTM), 93% (AE_LSTM), and 96% (CNN) are demonstrated by the results. CNN is the most effective of these, although AE_LSTM's use of compressed features offers promise for improvement. In addition to opening the door for the integration of more sophisticated models and characteristics, this study offers a scalable, real-time detection mechanism.