A Hardware Trojan Detection Method Based on SE-DenseNet-BC(SDBC) under GASF Preprocessing

Abstract Detecting hardware Trojans (HTs) using side-channel data has been a focus since 2007, with machine learning methods commonly used for detection. However, traditional methods now fall short compared to deep learning techniques. This paper proposes a new architecture, the SE-DenseNet-BC (SDBC) network, designed to detect HTs more efficiently and accurately. Inspired by the DenseBlock network with an added attention mechanism, the architecture first preprocesses side-channel data using the Gramian Angle Summation Field (GASF) technique, converting time-series data into feature-rich RGB images. Features are then extracted through a convolutional layer, with an attention module highlighting relevant data aspects. The processed feature maps are passed through the DenseBlock and attention modules to enhance classification. A fully connected layer performs the final classification. We evaluate the approach using AES family HTs from TrustHub, achieving up to 98.7% accuracy. The results show superior classification performance compared to existing deep learning methods, significantly contributing to chip security.