Error-Resilient PUF-Based Authentication on IoT Edge Devices Using Machine Learning
Nikhil Joshi, Nikolaos Athanasios Anagnostopoulos, Nico Mexis, Shekoufeh Neisarian, Jiayi Chang, Owen Millwood, Tolga Arul, Stefan Katzenbeisser, Elif Bilge Kavun · 2025
The security of Internet-of-Things (IoT) devices has become a pressing issue due to their interconnections and increasing number. Such devices are often resource-constrained and lack dedicated security mechanisms. Nevertheless, such a device could be secured by leveraging on-device Static Random Access Memory (SRAM) as a memory-based Physical Unclonable Function (PUF), which is cost-effective and requires no hardware modification. In this work, we implement and test an error-resilient authentication protocol using Machine Learning (ML) with a Convolutional Neural Network (CNN) architecture which can run on edge devices and authenticate intact, noisy, and corrupted SRAM PUF responses. The protocol runs inference on the EfficientNet-Lite and MobileNet models and the authentication is performed based on the confidence score for the classification of responses (images created from the SRAM PUF responses) received from these models.