Privacy-Preserving Data and Model Sharing for EDA via Inference Control and DP-SGD

In the semiconductor field, the lack of cutting-edge datasets hinders the advancement of machine learning (ML) applications in electronic design automation (EDA). While real-world datasets could significantly improve the soundness and impact of model development in academic research, industry remains reluctant to share data due to security, competitive, and regulatory concerns. This paper proposes two privacy-preserving scenarios for enabling secure data and model sharing in EDA: a feature obfuscation method based on k-anonymity to prevent inference attacks, and a transfer learning approach combined with differentially private stochastic gradient descent (DP-SGD) to support model sharing with quantifiable privacy guarantees. Through simulated industrial-academic collaboration experiments, we demonstrate that both approaches achieve a favorable trade-off between privacy and model utility in tasks such as congestion prediction and Power-Performance-Area(PPA) prediction, thereby providing a possibility of secure and practical knowledge transfer across organizational boundaries.