GNNs Learn Structural Manipulability in Gate-Level Netlists

Gate-level netlists, fundamental components in integrated circuits, possess inherent structural properties that dictate how signals propagate, independent of their functional behavior. This paper introduces a novel topology-driven structural manipulability score. This score quantifies the flexibility of individual nodes within a netlist by considering factors like path participation, k-core embedding, symmetry, and centrality. The researchers modeled netlists as directed graphs and applied Graph Neural Networks (GNNs) to learn this newly defined score through node-level regression. Experiments conducted on standard benchmarks, such as ISCAS85 and EPFL, showed that various GNN architectures could effectively approximate this metric, with hierarchical models demonstrating the most consistent performance across different circuits. An illustrative case study involving Trojan-injected circuits revealed that this topology-based scoring method could identify statistically distinct structural patterns. This indicates that understanding structural manipulability through GNNs offers complementary insights into circuit analysis, potentially aiding in design verification and security assessment.