PUF-based Lightweight Authentication for Binarized Neural Networks
Gokulnath Rajendran, Suman Deb, Anupam Chattopadhyay · 2024
Rapid adoption of Neural Networks (NN) for various applications have given rise to concerns about protecting the ownership of the Intellectual Property (IP). Recent proposals in that direction include techniques ranging from locking of NN to performing computation in encrypted domain. These techniques either suffer from high performance overhead or lack of concrete guarantee of user authentication. Moreover, key management for large-scale deployment always remains a challenge. In this work, we propose a lightweight authentication of Binarized Neural Networks (BNN). The key management is done using Physical Unclonable Functions (PUF), presenting a compact solution. The entire design is realized using RRAM crossbars, a platform of choice in recent times due to highly efficient NN processing. Through formal analysis and experimental evaluations, we corroborate the security claims (${\mathcal{O}}\left(2^{n}\right)$, where n is key size) as well as low power overhead (<1 %).