Nonlinear Ferroelectric Diode Offers High Security in Accurate and Energy-Efficient Compute-in-Memory

Neural network models implemented with compute-in-memory (CIM) architectures are vulnerable to readout attacks and reverse engineering. Existing encryption methods based on linear devices typically involve complex circuits, poor scalability, and high-power consumption. In this work, we propose a secure CIM architecture leveraging the intrinsic nonlinearity of ferroelectric diodes (FeDs) to enable encryption directly at the device level. A transistor-free, multilayer encryption scheme is embedded within the FeD array, incorporating XOR-based input–weight co-encryption, unclonable input obfuscation, and β-sensitive linearization. These mechanisms ensure that only correct keys yield accurate matrix-vector multiplication (MVM), while incorrect keys lead to masked and distorted outputs. Physical measurements on a 1.6 kb FeD crosspoint array show that inference accuracy drops from 86.33% to 18.04% after encryption, while decryption successfully restores it without any loss. This device-level approach is compatible with other nonlinear memories, offering a scalable and energy-efficient hardware security solution for CIM.