UN-SPLIT: Attacking Split Manufacturing Using Link Prediction in Graph Neural Networks.

We explore a new angle for attacking split manufacturing aside from relying only on physical design hints. By learning on the structure, composition, and the front-end-of-line (FEOL) interconnectivity of gates in a given design (or design library/dataset), along with key hints from physical design, we obtain a model that can predict the missing back-end-of-line (BEOL) connections. We formulate this as a link-prediction problem and solve it using a graph neural network (GNN). Furthermore, we utilize post-processing techniques that analyze the GNN predictions and apply common domain knowledge to further enhance the accuracy of our attack methodology.