Resumen
Logic locking is a technique that can help hinder reverse-engineering-based attacks in the IC supply chain from untrusted foundries or end-users. In 2015, the Boolean Satisfiability (SAT) attack was introduced. Although the SAT attack is effective in deobfuscating a wide range of logic locking schemes, its execution time varies widely from a few seconds to months. Previous research has shown that Graph Convolutional Networks (GCN) may be used to estimate this deobfuscation time for locked circuits with varied key sizes. In this paper, we explore whether GCN models truly understand/capture the structural/functional sources of deobfuscation hardness. In order to tackle this, we generate different curated training datasets: traditional ISCAS benchmark circuits locked with varying key sizes, as well as an important novel class of synthetic benchmarks: Substitution-Permutation Networks (SPN), which are circuit structures used to produce the most secure and efficient keyed-functions used today: block-ciphers. We then test whether a GCN trained on a traditional benchmark can predict the simple fact that a deeper SPN is superior to a wide SPN of the same size. We find that surprisingly the GCN model fails at this. We propose to overcome this limitation by proposing a set of circuit features motivated by block-cipher design principles. These features can be used as stand-alone or combined with GCN models to provide deeper topological cues than what GCNs can access.