Differential Emulatable PUF Design against Machine Learning Attacks for Hardware Security

Physical unclonable functions (PUFs) are emerged to provide lightweight hardware security primitives, by generating device-specific responses as physical fingerprints. Emulatable PUFs use emulatable models to generate secrete reference information, eliminating the need of databases to store such information. Emulatable PUFs further improves the security by providing execution-simulation time gap. Meanwhile, machine learning (ML) attacks have become one of the greatest concerns to produce software counterfeits for PUFs. This paper investigates a typical emulatable PUF, differential emulatable PUF (DEPUF). We demonstrate a multilayer perceptron-based model, to show DEPUF is potentially susceptible to ML attacks. We investigate how structural configurations of DEPUFs can affect their resilience against ML attacks. The results indicate the resilience can be most significantly improved by increasing the width of a DEPUF.