Understanding and Improving Fast Adversarial Training

Abstract

A recent line of work focused on making adversarial training computationally efficient for deep learning models. In particular, Wong et al. (2020) showed that ${\ell }_{\infty }$-adversarial training with fast gradient sign method (FGSM) can fail due to a phenomenon called "catastrophic overfitting", when the model quickly loses its robustness over a single epoch of training. We show that adding a random step to FGSM, as proposed in Wong et al. (2020), does not prevent catastrophic overfitting, and that randomness is not important per se -- its main role being simply to reduce the magnitude of the perturbation. Moreover, we show that catastrophic overfitting is not inherent to deep and overparametrized networks, but can occur in a single-layer convolutional network with a few filters. In an extreme case, even a single filter can make the network highly non-linear locally, which is the main reason why FGSM training fails. Based on this observation, we propose a new regularization method, GradAlign, that prevents catastrophic overfitting by explicitly maximizing the gradient alignment inside the perturbation set and improves the quality of the FGSM solution. As a result, GradAlign allows to successfully apply FGSM training also for larger ${\ell }_{\infty }$-perturbations and reduce the gap to multi-step adversarial training. The code of our experiments is available at https://github.com/tml-epfl/understanding-fast-adv-training.

arXiv commentThe camera-ready version (accepted at NeurIPS 2020)

References in corpus (28)

• Deep Learning in Neural Networks: An Overview
• YOLOv4: Optimal Speed and Accuracy of Object Detection
• Ensemble Adversarial Training: Attacks and Defenses
• Wild Patterns: Ten Years After the Rise of Adversarial Machine Learning
• Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
• Mixed Precision Training
• Theoretically Principled Trade-off between Robustness and Accuracy
• Provable defenses against adversarial examples via the convex outer adversarial polytope
• Towards Deep Neural Network Architectures Robust to Adversarial Examples
• Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks
• Certified Defenses against Adversarial Examples
• Fast is better than free: Revisiting adversarial training
• Using Pre-Training Can Improve Model Robustness and Uncertainty
• Understanding Adversarial Training: Increasing Local Stability of Neural Nets through Robust Optimization
• Robustness May Be at Odds with Accuracy
• Evaluating Robustness of Neural Networks with Mixed Integer Programming
• FreeLB: Enhanced Adversarial Training for Natural Language Understanding
• Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing their Input Gradients
• Formal Guarantees on the Robustness of a Classifier against Adversarial Manipulation
• Adversarial Robustness through Local Linearization
• On the Convergence and Robustness of Adversarial Training
• Provably Minimally-Distorted Adversarial Examples
• Image Synthesis with a Single (Robust) Classifier
• First-order Adversarial Vulnerability of Neural Networks and Input Dimension
• An Alternative Surrogate Loss for PGD-based Adversarial Testing
• SGD on Neural Networks Learns Functions of Increasing Complexity
• Overfitting in adversarially robust deep learning
• A Closer Look at Double Backpropagation

Cited by in corpus (10)

• Survey on Adversarial Attack and Defense for Medical Image Analysis: Methods and Challenges
• Recent Advances in Adversarial Training for Adversarial Robustness
• On the effectiveness of adversarial training against common corruptions
• SoK: Machine Learning Governance
• Understanding Catastrophic Overfitting in Adversarial Training
• ZeroGrad : Mitigating and Explaining Catastrophic Overfitting in FGSM Adversarial Training
• Deep Repulsive Prototypes for Adversarial Robustness
• Robust Single-step Adversarial Training with Regularizer
• Lagrangian Objective Function Leads to Improved Unforeseen Attack Generalization in Adversarial Training
• Local Linearity and Double Descent in Catastrophic Overfitting