A Monolithic Reconfigurable RRAM CIM Array Integrating PUF, TRNG, and a Lightweight Block Cipher for Secure Edge AI
Lists
Only you see this
Abstract
AI edge devices urgently need on-chip security that is both area- and energy-efficient, yet current designs still separate security, compute, and storage blocks, inflating silicon area and power consumption. We designed SCIM, an RRAM-based secure compute-in-memory chip that fully integrates the entire cryptographic security algorithm suite (including cipher module, PUF, and TRNG), leveraging the energy-efficient reconfigurable RRAM array. SCIM consists of five functions: (i) ISO/IEC-compliant PRESENT cipher, (ii) reconfigurable r-PUF, (iii) a TRNG with entropy throughput of 357 Mbps, (iv) one-time pad (OTP)-style encryption/decryption, and (v) vector–matrix multiplication. Consolidating these primitives in one array shrinks area by 69.1%. The r-PUF achieves a Hamming distance of 49.87% between chips with a concealment error of 20.12%. In-array PRESENT reduces latency by 25.9% and energy by 92.0% versus conventional designs. Moreover, XOR encryption of CIFAR-100 classification heads cuts accuracy by 78.23% (coarse) and 82.04% (fine), but completely restores performance after licensed decryption. Collectively, SCIM delivers a low-area, multifunctional, and energy-efficient hardware root of trust for next-generation edge AI applications.
BibTeX
@misc{wang2025monolithic,
author = {Songqi Wang and Shuang Liang and Shaonan Wu and Zhiqi Yang and Jichang Yang and Xinyuan Zhang and Yi Li and Yong Zhang and Zi Wang and Aoyang Zhang and Han Wang},
title = {{A Monolithic Reconfigurable RRAM CIM Array Integrating PUF, TRNG, and a Lightweight Block Cipher for Secure Edge AI}},
year = {2025},
doi = {10.1109/iedm50572.2025.11353751},
}Loading related content...