ML for Attack and Defense of PUFs: Current Status and Future Directions

The integration of IoT devices is becoming increasingly inevitable in the development of next-generation systems and applications. Due to such a wide adoption, IoT devices handle large quantities of private and sensitive data, and operate safety-critical systems. As such, failure to comply with security requirements would prove to be catastrophic. However, the resource-constrained nature of IoT devices is a fundamental limitation in designing their security features. To tackle the problem of implementing lightweight security functionalities that enable trusted communications, Physical Unclonable Functions (PUFs) have been proposed. Exploiting the manufacturing variations of Integrated Circuits (ICs), these primitives aim to give devices a unique identifier that no attacker can violate or clone. That said, in the past decade many studies have shown the great threat that Machine Learning (ML) poses to the security of Physical Unclonable Functions. In this paper, we provide an up-to-date situation of this field of research, as well as our current work and future directions.