Skip to content
DSML.KZ
Publications

Talk · International Conference on Machine Intelligence and Smart Innovation · May 10, 2025

Adversarial Robustness and Transferability on Stochastic Resistive Accelerators

Kamilya Smagulova, Mohammed E. Fouda, Ahmed M. Eltawil

Kamilya Smagulova
Open via DOI

Abstract

Adversarial attacks compromise an integrity of machine learning models, posing risks to critical applications. To counter these threats, this study leverages the stochastic nature of real-world data and non-volatile memory (NVM) devices in crossbar arrays through the Neural Sampling Machine (NSM) model, a new type of stochastic neural network. We assess NSM-based image classifiers on MNIST and CIFAR10 datasets against adversarial attacks using the Adversarial Robustness Toolbox (ART). The gradient-based FGSM (Fast Gradient Sign Method) and C&W$L_{\infty}$(Carlini & Wagner$L_{\infty}$) attacks showed low success rates, with minimal impact on classification accuracy. NSM's weight normalization significantly improved resilience against C&W$L_{2}$attacks, boosting accuracy from 0.48% to over 93% for MNIST, and from 3.05% to over 63% for CIFAR10. Additionally, NSM exhibited superior robustness over deterministic models against iterative attacks such as HopSkipJump, SquareAttack and AutoAttack, highlighting its potential to enhance machine learning security.

Adversarial Robustness in Machine LearningSecurity and Verification in ComputingPhysical Unclonable Functions (PUFs) and Hardware Securityadversarial robustnessadversarial transferabilitystochastic neural networks

More by Kamilya Smagulova