NETLAM: An Automated LLM Framework to Generate and Evaluate Stealthy Hardware Trojans

Tishya Sarma Sarkar, Kislay Arya, Siddhartha Chowdhury, Upasana Mandal, Shubhi Shukla, Sarani Bhattacharya, Debdeep Mukhopadhyay

Abstract: Securing externally sourced hardware designs is essential to prevent adversaries from embedding hardware Trojans. Trojans are stealthy modifications that leak data or create backdoors. Existing benchmarks like Trust-Hub provide only a limited set of Trojans (106), while the possibilities are virtually infinite. To address this, we propose NETLAM, a comprehensive framework utilizing multiple LLM-based tools to generate previously undiscovered Trojans not included in Trust-Hub. The first tool converts hardware netlists into Directed Acyclic Graphs (DAGs) to identify vulnerable nets and components in digital designs. Using these insights, the second tool generates stealthy Trojan-infected versions of the original design. To evaluate the stealthiness of these Trojans, we use an LLM-based equivalence checker, where stealthier Trojans pass equivalence checks while others are detected. We evaluate NETLAM using the AES dataset from Trust-Hub consisting of 28 Trojans. We identified 5 new Trojans, with high Common Vulnerability Scoring System (CVSS) scores, demonstrating their stealthiness. To prove the efficacy of the NETLAM generated Trojans, we further utilize an open-source formal equivalence checker to perform a functional equivalence check between the golden and the NETLAM generated Trojan-infected circuits. All of the suggested Trojans pass the formal equivalence check. However, the same Trojan-infested circuits fail in the NETLAM equivalence test, thus validating the effectiveness of our proposed framework. We show that LLMs and Generative AI models, such as GPT-4o and Gemini, can enhance Trojan detection by using semantic and probabilistic analysis rather than strict logical equivalence (GitHub Repository: https://github.com/shubhishukla10/NETLAM).

External IDs:dblp:conf/acns/SarkarACMSBM25