Maximum-Entropy Adversarial Data Augmentation for Improved Generalization and Robustness

Abstract

Adversarial data augmentation has shown promise for training robust deep neural networks against unforeseen data shifts or corruptions. However, it is difficult to define heuristics to generate effective fictitious target distributions containing "hard" adversarial perturbations that are largely different from the source distribution. In this paper, we propose a novel and effective regularization term for adversarial data augmentation. We theoretically derive it from the information bottleneck principle, which results in a maximum-entropy formulation. Intuitively, this regularization term encourages perturbing the underlying source distribution to enlarge predictive uncertainty of the current model, so that the generated "hard" adversarial perturbations can improve the model robustness during training. Experimental results on three standard benchmarks demonstrate that our method consistently outperforms the existing state of the art by a statistically significant margin.

arXiv commentAccepted to NeurIPS 2020. Code is available at https://github.com/garyzhao/ME-ADA

References in corpus (13)

• Auto-Encoding Variational Bayes
• What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?
• Improved Regularization of Convolutional Neural Networks with Cutout
• The information bottleneck method
• Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift
• Bayesian Convolutional Neural Networks with Bernoulli Approximate Variational Inference
• Domain Separation Networks
• On Mutual Information Maximization for Representation Learning
• Learning Robust Global Representations by Penalizing Local Predictive Power
• Learning Representations for Neural Network-Based Classification Using the Information Bottleneck Principle
• Learning Robust Representations by Projecting Superficial Statistics Out
• Testing Robustness Against Unforeseen Adversaries
• Uncertainty in the Variational Information Bottleneck

Cited by in corpus (4)

• InfoAT: Improving Adversarial Training Using the Information Bottleneck Principle
• Adversarial Parameter Defense by Multi-Step Risk Minimization
• Adversarial AutoMixup
• Learning Less Generalizable Patterns with an Asymmetrically Trained Double Classifier for Better Test-Time Adaptation