Noisy-Defense Variational Auto-Encoder (ND-VAE): an Adversarial Defense Framework to Eliminate Adversarial Attacks

Published In

2023 Fifth International Conference on Transdisciplinary AI (transai)

Document Type

Citation

Publication Date

2023

Abstract

This paper presents a robust adversarial defense mechanism, Noisy-Defense Variational Auto-Encoder (ND-VAE), that combines the strengths of Nouveau VAE (NVAE) and Defense-VAE to effectively eliminate adversarial attacks from contaminated images. The ND-VAE, trained using noisy images, not only removes adversarial perturbations but also preserves the image details, thereby lowering adversarial training costs. By utilizing advanced NVAE architectures and incorporating a noise filter, the defense system efficiently mitigates both previously seen and unseen adversarial attacks. Our evaluations on both MNIST and Fashion-MNIST datasets confirm the high efficiency of ND-VAE, including better zero-shot performance, demonstrating a balanced approach to model expressivity and noise resistance in image classifiers.11Code available at https://github.com/shayan223/ND-VAE

DOI

10.1109/TransAI60598.2023.00018

Persistent Identifier

https://archives.pdx.edu/ds/psu/41277

Publisher

IEEE

Share

COinS