(A) trainable physically unclonable function with unassailability against deep learning attacks using memristor array

Abstract

The proliferation of edge devices necessitates new needs for privacy protection and chip authentication primitives. Due to its natural randomization and stochastic features, memristors are an appealing entropy source for implementing hardware-based security primitives. We propose a trainable encryption system, an unique physically unclonable function (PUF) architecture that utilizes deep learning techniques, with the deployment of memristors as one of many technologies that match basic requirements. Because the system does not rely on the precision of write operations, the suggested architecture coupled with the memristor array is simple to develop. To consider a non-differentiable module during training, we provide a novel concept of loss termed PUF loss. Update of weights using the loss function results in the best performance. We demonstrate that our design yields a near-ideal average value of 50 percent for security criteria such as uniformity, diffuseness, and uniqueness. This indicates that through training with PUF loss, our design meets the realistic quality criteria for security primitives. Additionally, we demonstrate that the resultant response is impervious to deep learning-based modeling attacks, as illustrated by the near-50 percent prediction model accuracy.