Adversarial Diversity and Hard Positive Generation

TL;DR

This paper introduces a new method for generating diverse and semantically meaningful adversarial examples to improve neural network robustness, using a novel hot/cold approach and a perceptual similarity score.

Contribution

It presents a novel adversarial example generation technique that produces diverse, meaningful perturbations and demonstrates improved robustness through fine-tuning with these examples.

Findings

Diverse adversarial images increase model robustness.

The hot/cold approach yields semantically meaningful perturbations.

Fine-tuning with hard positives improves network robustness.

Abstract

State-of-the-art deep neural networks suffer from a fundamental problem - they misclassify adversarial examples formed by applying small perturbations to inputs. In this paper, we present a new psychometric perceptual adversarial similarity score (PASS) measure for quantifying adversarial images, introduce the notion of hard positive generation, and use a diverse set of adversarial perturbations - not just the closest ones - for data augmentation. We introduce a novel hot/cold approach for adversarial example generation, which provides multiple possible adversarial perturbations for every single image. The perturbations generated by our novel approach often correspond to semantically meaningful image structures, and allow greater flexibility to scale perturbation-amplitudes, which yields an increased diversity of adversarial images. We present adversarial images on several network topologies and datasets, including LeNet on the MNIST dataset, and GoogLeNet and ResidualNet on the ImageNet dataset. Finally, we demonstrate on LeNet and GoogLeNet that fine-tuning with a diverse set of hard positives improves the robustness of these networks compared to training with prior methods of generating adversarial images.