Machine learning analysis on logic locked circuits

Abstract

To reduce the design effort and cost, now many semiconductor companies are fabless. During the outsource fabrication and test process, logic locking is widely used to protect their intellectual Property (IP) from untrustworthy access. Logic locking conceals the designed functionality by adding key inputs together with additional gates to the original circuit. Without the correct key inputs, attackers can not reveal the encrypted design. In the past decades, several attack methods were proposed to test the vulnerability of existing logic locking methods. Meanwhile, to counter these attack methods, novel defences technology were also developed. In recent years, Machine learning based attacking methods became prominent. Compared with traditional attack methods, oracle (an unlocking circuit) is not needed in machine-learning-based attacks. But their accuracy is unsatisfactory. In this paper, different logic locking attack methods will be introduced and compared. Graph Neural Network (GNN) will be used to decode the locked circuits. The netlists are naturally converted in to the neural graph, and the keys are derived from classification task of neural networks. The relationship between the accuracy and circuit size will also be summarized.