GAE4HT: Detecting Hardware Trojans with Graph Autoencoder-Trained on Golden Model Data Flow Graphs

Hardware Trojans (HTs) are malicious circuits covertly embedded in hardware, similar to software backdoors, that can cause malfunctions or leak confidential information when activated. Detecting and localizing these HTs is critical, especially in sensitive systems like weaponry, where they pose significant security threats. Traditional detection methods often rely on side-channel analysis, which requires specialized equipment and expertise, making them costly and limited to post-manufacturing stages. Localization is crucial because merely detecting an HT does not allow for effective neutralization. Pinpointing the exact location helps in assessing the impact, facilitating targeted removal, and avoiding costly hardware redesigns. In critical applications like military systems, this minimizes the risk of security breaches that could compromise the entire system. This paper proposes a novel method for HT detection and localization using a Graph Autoencoder (GAE). Unlike existing techniques, our approach utilizes the structural information in Data Flow Graphs (DFGs) to identify HTs without the need for labeled datasets or manual feature engineering. By training solely on the Golden Model (GM) of the hardware design, the GAE captures the circuit's normal structural patterns. Our method analyzes the reconstruction error of each node in the GAE to not only detect the presence of HTs but also accurately pinpoint their insertion points within the hardware design. Experimental results show that our method outperforms existing approaches across various benchmarks, achieving high detection accuracy. It also offers practical advantages by avoiding the need for additional equipment or HT data during training, making it a feasible solution for early HT detection and mitigation in hardware systems.