Adversarial attacks on an optical neural network

TL;DR

This paper demonstrates that optical neural networks are vulnerable to adversarial attacks, showing that small, imperceptible image perturbations can cause misclassification, highlighting security concerns in optical machine learning.

Contribution

First to propose an adversarial attack scheme on optical neural networks, revealing their susceptibility to such attacks.

Findings

Adversarial attacks cause misclassification in ONNs with minimal visual differences

Optical neural networks are vulnerable to adversarial perturbations

The attack scheme effectively deceives the ONN classifier

Abstract

Adversarial attacks have been extensively investigated for machine learning systems including deep learning in the digital domain. However, the adversarial attacks on optical neural networks (ONN) have been seldom considered previously. In this work, we first construct an accurate image classifier with an ONN using a mesh of interconnected Mach-Zehnder interferometers (MZI). Then a corresponding adversarial attack scheme is proposed for the first time. The attacked images are visually very similar to the original ones but the ONN system becomes malfunctioned and generates wrong classification results in most time. The results indicate that adversarial attack is also a significant issue for optical machine learning systems.