Improving adversarial robustness by learning shared information

doi:10.1016/j.patcog.2022.109054

Improving adversarial robustness by learning shared information

2023-02-01

13

PRE

AI

X

Xi Yu

N

Niklas Smedemark-Margulies

DOI：10.1016/j.patcog.2022.109054

Original

Original request for help

Save

Abstract

En 中文

We consider the problem of improving the adversarial robustness of neural networks while retaining nat-ural accuracy. Motivated by the multi-view information bottleneck formalism, we seek to learn a repre-sentation that captures the shared information between clean samples and their corresponding adversar-ial samples while discarding these samples' view-specific information. We show that this approach leads to a novel multi-objective loss function, and we provide mathematical motivation for its components to-wards improving the robust vs. natural accuracy tradeoff. We demonstrate enhanced tradeoff compared to current state-of-the-art methods with extensive evaluation on various benchmark image datasets and architectures. Ablation studies indicate that learning shared representations is key to improving perfor-mance.(c) 2022 Elsevier Ltd. All rights reserved.

Keywords:

Adversarial robustness

Information bottleneck

Multi-view learning

Shared information

Improving adversarial robustness by learning shared information

Abstract

Journal

Organization