An Ensemble Approach Towards Adversarial Robustness

TL;DR

This paper introduces an ensemble method that improves adversarial robustness by dividing complex tasks into simpler ones and aggregating their outputs, achieving state-of-the-art results on MNIST and Fashion-MNIST without adversarial training.

Contribution

It proposes a novel ensemble approach with fractal divide and aggregation techniques that enhance adversarial robustness and provable guarantees for image classifiers.

Findings

Achieved 99% natural accuracy on MNIST with 70% robustness

Achieved 90% natural accuracy on Fashion-MNIST with 54.5% robustness

Set new state-of-the-art results on binary label-pairs for robustness

Abstract

It is a known phenomenon that adversarial robustness comes at a cost to natural accuracy. To improve this trade-off, this paper proposes an ensemble approach that divides a complex robust-classification task into simpler subtasks. Specifically, fractal divide derives multiple training sets from the training data, and fractal aggregation combines inference outputs from multiple classifiers that are trained on those sets. The resulting ensemble classifiers have a unique property that ensures robustness for an input if certain don't-care conditions are met. The new techniques are evaluated on MNIST and Fashion-MNIST, with no adversarial training. The MNIST classifier has 99% natural accuracy, 70% measured robustness and 36.9% provable robustness, within L2 distance of 2. The Fashion-MNIST classifier has 90% natural accuracy, 54.5% measured robustness and 28.2% provable robustness, within L2 distance of 1.5. Both results are new state of the art, and we also present new state-of-the-art binary results on challenging label-pairs.