Can We Trust Machine Learning for Electronic Design Automation?

Machine learning (ML) techniques, especially deep learning (DL), have been playing a vital role in developing electronic design automation (EDA) tools across the chip design flow. DL for EDA demonstrates comparable quality, faster turnaround time, and better generalization than traditional methods. However, DL techniques expose intrinsic vulnerabilities to adversarial perturbation attacks and backdooring attacks. Despite the expansive application and great promise of DL use in the computer-aided design (CAD) flow, compromised DL-based EDA tools can propagate undetected design errors, result in manufacturability issues, and eventually reduce downstream integrated circuit (IC) yields. With lithographic hotspot detection as a case study, this paper highlights current challenges and recent progress for securing and robustifying lithographic hotspot detectors and sheds insights into the future directions in enhancing the security and robustness of DL for EDA.