Correlation-Based Robust Authentication (Cobra) Using Helper Data Only

Abstract Physical unclonable function (PUF)-based authentication protocols have been proposed as a strong challenge-response form of authentication for internet of things (IoT) and embedded applications. A special class of so called strong PUFs are best suited for authentication because they are able to generate an exponential number of challenge-response-pairs (CRPs). However, strong PUFs must also be resilient to model-building attacks. Model-building utilizes machine learning algorithms and a small set of CRPs to build a model that is able to predict the responses of a fielded chip, thereby compromising the security of chip-server interactions. In this paper, response bitstrings are eliminated in the message exchanges between chips and the server during authentication, and therefore, it is no longer possible to carry out model-building attacks in the traditional manner. Instead, the chip transmits a Helper Data bitstring to the server and this information is used for authentication instead. The server constructs Helper Data bitstrings using enrollment data that it stores for all valid chips in a secure database and computes correlation coefficients (CCs) between the chip’s Helper Data bitstring and each of the server-generated Helper Data bitstrings. The server authenticates (and identifies) the chip if a CC is found that exceeds a threshold, which is determined during characterization. The technique is demonstrated using data from a set of 500 Xilinx Zynq 7020 FPGAs, subjected to industrial-level temperature and voltage variations. View Full-Tex