Neural-Network-Based Modeling Attacks on XOR Arbiter PUFs Revisited
Published on Cryptology ePrint Archive
By revisiting recent neural-network based modeling attacks on XOR Arbiter PUFs from the literature, we show that XOR Arbiter PUFs and Interpose PUFs can be attacked faster, up to larger security parameters, and with orders of magnitude fewer challenge-response pairs than previously known. To support our claim, we discuss the differences and similarities of recently proposed modeling attacks and offer a fair comparison of the performance of these attacks by implementing all of them using the popular machine learning framework Keras and comparing their performance against the well-studied Logistic Regression attack. Our findings show that neural-network-based modeling attacks have the potential to outperform traditional modeling attacks on PUFs and must hence become part of the standard toolbox for PUF security analysis; the code and discussion in this paper can serve as a basis for the extension of our results to PUF designs beyond the scope of this work.