Towards Robust Stacked Capsule Autoencoder with Hybrid Adversarial Training

TL;DR

This paper identifies vulnerabilities in the stacked capsule autoencoder (SCAE) to adversarial attacks and proposes Hybrid Adversarial Training (HAT) to improve its robustness, achieving over 82% accuracy under attack.

Contribution

It introduces a novel evasion attack on SCAE and proposes Hybrid Adversarial Training (HAT) as a defense, enhancing robustness against adversarial perturbations.

Findings

Refined SCAE achieves 82.14% accuracy under attack.

Proposed HAT improves robustness of SCAE.

Adversarial perturbations effectively fool SCAE models.

Abstract

Capsule networks (CapsNets) are new neural networks that classify images based on the spatial relationships of features. By analyzing the pose of features and their relative positions, it is more capable to recognize images after affine transformation. The stacked capsule autoencoder (SCAE) is a state-of-the-art CapsNet, and achieved unsupervised classification of CapsNets for the first time. However, the security vulnerabilities and the robustness of the SCAE has rarely been explored. In this paper, we propose an evasion attack against SCAE, where the attacker can generate adversarial perturbations based on reducing the contribution of the object capsules in SCAE related to the original category of the image. The adversarial perturbations are then applied to the original images, and the perturbed images will be misclassified. Furthermore, we propose a defense method called Hybrid Adversarial Training (HAT) against such evasion attacks. HAT makes use of adversarial training and adversarial distillation to achieve better robustness and stability. We evaluate the defense method and the experimental results show that the refined SCAE model can achieve 82.14% classification accuracy under evasion attack. The source code is available at https://github.com/FrostbiteXSW/SCAE_Defense.