DRAM-Based Authentication Using Deep Convolutional Neural Networks
Michael Yue, Nima Karimian, Wei Yan, Nikolaos Athanasios Anagnostopoulos, Fatemeh Tehranipoor
Abstract
Authentication is the act of proving that an integrated circuit (IC) is not counterfeit. One application of a physical unclonable function (PUF) circuit is to authenticate the identity of the chip using raw bits of the memory. However, several previous works present machine learning-based modeling attacks on PUFs. To alleviate this issue, we propose a novel authentication scheme involving unique DRAM power-up values using a deep convolutional neural network (CNN). This methodology eliminates the need for PUFs and can authenticate DRAM technology accurately with a neural network. Our approach converts raw power-up sequence data from DRAM cells into a two-dimensional (2D) format to generate a DRAM image structure. This makes it harder for an adversary to use machine learning since there is no PUF to exploit the weaknesses. Then, we apply deep CNN to DRAM images to extract unique features from each chip and classify them for authentication. Our method “DRAMNet” achieves 98.84% accuracy and 98.73% precision. The proposed technique has the advantage of a faster authentication while eliminating the need for costly error correction mechanisms and CRPs. To the best of our knowledge, this is the first method to authenticate ICs using DRAM and CNN.