Discerning the Limitations of GNN-based Attacks on Logic Locking
DescriptionMachine learning (ML)-based attacks have revealed the possibility of utilizing neural networks to break locked circuits without a need for functional chips or oracles. Among ML approaches, GNN-based attacks are the most powerful tools that attackers can utilize as they exploit graph structures which is the very nature of the circuits' netlist. Although promising, GNNs have some limitations in attacking logic locking. In this paper, we investigate the limitations of the state-of-the-art GNN-based attacks against logic locking. We show that by utilizing these limitations in the key gate selection process, we can drastically decrease the accuracy of these attacks.
TimeThursday, July 13th4:40pm - 4:55pm PDT
Location3002, 3rd Floor
Hardware Security: Attack and Defense