ConVERTS: Contrastively Learning Structurally InVariant Netlist Representations

Graph neural network (GNN)-based representations of hardware designs are used in electronic design automation (EDA) tasks like logic synthesis, verification, and hardware security. While promising, state-of-the-art methods are supervised and require target labels and/or need different behavioral register transfer level (RTL) codes of the same function as training data to generalize. We propose ConVERTS, a self-supervised netlist contrastive learning method that generalizes well using one-shot RTL of a design. We demonstrate the effectiveness of ConVERTS on two use-cases: (1) netlist classification, and (2) Recovering functionality of obfuscated designs.