Fixing Data Augmentation to Improve Adversarial Robustness

Abstract

Adversarial training suffers from robust overfitting, a phenomenon where the robust test accuracy starts to decrease during training. In this paper, we focus on both heuristics-driven and data-driven augmentations as a means to reduce robust overfitting. First, we demonstrate that, contrary to previous findings, when combined with model weight averaging, data augmentation can significantly boost robust accuracy. Second, we explore how state-of-the-art generative models can be leveraged to artificially increase the size of the training set and further improve adversarial robustness. Finally, we evaluate our approach on CIFAR-10 against ${\ell }_{\infty }$ and ${\ell }_2$ norm-bounded perturbations of size $\epsilon =8/255$ and $\epsilon =128/255$, respectively. We show large absolute improvements of +7.06% and +5.88% in robust accuracy compared to previous state-of-the-art methods. In particular, against ${\ell }_{\infty }$ norm-bounded perturbations of size $\epsilon =8/255$, our model reaches 64.20% robust accuracy without using any external data, beating most prior works that use external data.

arXiv commentSince its original publication (2 Mar 2021), this paper has been accepted to NeurIPS 2021 as two separate and updated papers (Rebuffi et al., 2021; Gowal et al., 2021). The new papers improve results and clarity

References in corpus (31)

• Adam: A Method for Stochastic Optimization
• Large Scale GAN Training for High Fidelity Natural Image Synthesis
• Gaussian Error Linear Units (GELUs)
• Improved Regularization of Convolutional Neural Networks with Cutout
• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
• Ensemble Adversarial Training: Attacks and Defenses
• Theoretically Principled Trade-off between Robustness and Accuracy
• The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
• AutoAugment: Learning Augmentation Policies from Data
• CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features
• Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks
• GANSpace: Discovering Interpretable GAN Controls
• Adversarially Robust Generalization Requires More Data
• Using Pre-Training Can Improve Model Robustness and Uncertainty
• Adversarial Risk and the Dangers of Evaluating Against Weak Attacks
• Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs
• Adversarial Robustness through Local Linearization
• Improving Adversarial Robustness via Promoting Ensemble Diversity
• Uncovering the Limits of Adversarial Training against Norm-Bounded Adversarial Examples
• Classification Accuracy Score for Conditional Generative Models
• Minimally distorted Adversarial Examples with a Fast Adaptive Boundary Attack
• RobustBench: a standardized adversarial robustness benchmark
• Self-Adaptive Training: beyond Empirical Risk Minimization
• Adversarially Robust Generalization Just Requires More Unlabeled Data
• Controlling generative models with continuous factors of variations
• Ensemble Methods as a Defense to Adversarial Perturbations Against Deep Neural Networks
• An Alternative Surrogate Loss for PGD-based Adversarial Testing
• Logit Pairing Methods Can Fool Gradient-Based Attacks
• Data Augmentation Can Improve Robustness
• Learnable Boundary Guided Adversarial Training
• Strength in Numbers: Trading-off Robustness and Computation via Adversarially-Trained Ensembles

Cited by in corpus (18)

• How Deep Learning Sees the World: A Survey on Adversarial Attacks & Defenses
• RobustBench: a standardized adversarial robustness benchmark
• Unsolved Problems in ML Safety
• What You See is Not What the Network Infers: Detecting Adversarial Examples Based on Semantic Contradiction
• Post-breach Recovery: Protection against White-box Adversarial Examples for Leaked DNN Models
• Detecting Adversarial Examples Is (Nearly) As Hard As Classifying Them
• Adversarial Robustness through the Lens of Convolutional Filters
• HateProof: Are Hateful Meme Detection Systems really Robust?
• Robust Learning Meets Generative Models: Can Proxy Distributions Improve Adversarial Robustness?
• DiffDefense: Defending against Adversarial Attacks via Diffusion Models
• Certified Adversarial Defenses Meet Out-of-Distribution Corruptions: Benchmarking Robustness and Simple Baselines
• HO-FMN: Hyperparameter Optimization for Fast Minimum-Norm Attacks
• Score-Based Generative Classifiers
• Improving Fast Minimum-Norm Attacks with Hyperparameter Optimization
• Robustifying ${\ell }_{\infty }$ Adversarial Training to the Union of Perturbation Models
• Bridged Adversarial Training
• Noisy Feature Mixup
• HIRE-SNN: Harnessing the Inherent Robustness of Energy-Efficient Deep Spiking Neural Networks by Training with Crafted Input Noise