Securing Binarized Neural Networks via PUF-Based Key Management in Memristive Crossbar Arrays

Abstract

Binarized neural networks (BNNs) are a subset of deep neural networks proposed to consume less computational resources with a smaller energy budget. Recent studies showed that memristor-based in-memory computing architectures can be constructed to accelerate BNNs, with better performance compared to traditional CMOS technologies. The memristor nonvolatility utilized for in-memory computing poses a notable threat to theft attacks in the presence of adversaries with physical access. This motivates us to introduce two novel protection methodologies to safeguard the model parameters of BNNs in the memristive crossbar. We propose to take advantage of physical unclonable functions (PUFs), which can be implemented using memristor-based crossbars for protecting BNN. This feature provides superior security compared to the traditional stored-key-based schemes. We provide circuit-level hardware designs to implement our methodologies with negligible additional overhead compared to an unprotected design and detailed supporting analysis to validate our security claims.