One Crossbar, Two Functions: Analogue Image Obfuscation and Feature Extraction via Memristors

P. Janiszyn, A. Wąsiak-Maciejak, P. Sitarz, T. Matusiak (SEMIQA, Poland)

This work introduces a memristor-based hardware architecture for secure and efficient edge-to-cloud image transmission. The system leverages a memristor crossbar to simultaneously achieve signal obfuscation and device authentication through a single analog-domain computation. We employed a Python-based framework integrating IBM’s AIHWKit to model memristor process variation as a physical fingerprint, and MemTorch to simulate the filters and the entire circuit layout. The approach exploits two complementary sources of variability: cycle-to-cycle variation for dynamic obfuscation and device-to-device conductance spread to create a persistent hardware-intrinsic security mechanism. We further demonstrate that the crossbar operation serves as a feature extraction engine; specifically, we simulate a Sobel filter for edge detection that simultaneously obfuscates the raw image while delivering a recoverable edge feature map to authorized receivers. Simulations confirm robust obfuscation, strong diffusion, and reliable reconstruction, highlighting a promising route toward low-power, hardware-intrinsic security in next-generation edge imaging.