Machine Learning Attack Resistant Area-Efficient Reconfigurable Ising-PUF
Eslam Elmitwalli, Kai Ni, Selçuk Köse
Abstract
The Ising-physical unclonable function (PUF) is a recent PUF structure formed of a network of APUFs inspired by the Ising model. A large challenge–response pair (CRP) space with high resilience against machine learning modeling attacks can be attained due to the unique arrangement. These advantages, however, are achieved at the cost of a large area overhead. In this article, a reconfigurable Ising-PUF is introduced with several new design knobs to generate a much larger CRP space within a smaller area. A 25% increase in the number of challenge bits can be achieved for a design that occupies 30% of the conventional Ising-PUF area. With the proposed lightweight area-efficient design, up to 5.6 times lower area per CRP can be achieved compared to the existing design. Several improvements are proposed that leverage the large design space, enabling dynamic tradeoffs between the area and CRP pool with the proposed flexible customization of Ising-PUFs. A detailed analysis of this improved design space is explored for different parameters. The state-of-the-art machine learning modeling attacks are investigated, and the Ising-PUF structure is shown to be resilient.