Identification and Classification of Corrupted PUF Responses via Machine Learning
Reshmi Suragani, Emiliia Nazarenko, Nikolaos Athanasios Anagnostopoulos, Nico Mexis, Elif Bilge Kavun
Abstract
A Physical Unclonable Function (PUF) exploits de-vice manufacturing variations to extract a number of unique responses, each of which corresponds to a specific challenge, and use them for secret generation as well as authentication. However, these responses are often excessively noisy for authentication. In this study, in order to avoid using a complicated and costly Error Correction Code (ECC) in the context of a fuzzy extractor, we propose a Machine Learning (ML)-based classification technique that works accurately even in the presence of corrupted responses and we test it on an existing SRAM PUF dataset. In this approach, the SRAM power-up responses are transformed into 2D images in order to extract features from them for classification. With our proposed architecture, an accuracy of 97.34 % is achieved for “intact” noisy images. In addition, the presented ML model is successful in classifying responses even when a large part of them was corrupted. With our approach, the model can classify responses accurately with up to a 30 % loss in data. Finally, we also propose a proof-of-concept authentication technique using the relevant Convolutional Neural Network (CNN).