A Unified Game-Theoretic Interpretation of Adversarial Robustness

Abstract

This paper provides a unified view to explain different adversarial attacks and defense methods, i.e. the view of multi-order interactions between input variables of DNNs. Based on the multi-order interaction, we discover that adversarial attacks mainly affect high-order interactions to fool the DNN. Furthermore, we find that the robustness of adversarially trained DNNs comes from category-specific low-order interactions. Our findings provide a potential method to unify adversarial perturbations and robustness, which can explain the existing defense methods in a principle way. Besides, our findings also make a revision of previous inaccurate understanding of the shape bias of adversarially learned features.

References in corpus (41)

• Explaining and Harnessing Adversarial Examples
• Axiomatic Attribution for Deep Networks
• Improved Regularization of Convolutional Neural Networks with Cutout
• Wide Residual Networks
• ZOO: Zeroth Order Optimization based Black-box Attacks to Deep Neural Networks without Training Substitute Models
• Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples
• Theoretically Principled Trade-off between Robustness and Accuracy
• Delving into Transferable Adversarial Examples and Black-box Attacks
• Certified Adversarial Robustness via Randomized Smoothing
• Consistent Individualized Feature Attribution for Tree Ensembles
• Characterizing Adversarial Subspaces Using Local Intrinsic Dimensionality
• Formal Guarantees on the Robustness of a Classifier against Adversarial Manipulation
• Exploring the Landscape of Spatial Robustness
• Keeping the Bad Guys Out: Protecting and Vaccinating Deep Learning with JPEG Compression
• Understanding Global Feature Contributions With Additive Importance Measures
• On the Convergence and Robustness of Adversarial Training
• Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets
• Beyond Word Importance: Contextual Decomposition to Extract Interactions from LSTMs
• Houdini: Fooling Deep Structured Prediction Models
• Structured Adversarial Attack: Towards General Implementation and Better Interpretability
• Biologically inspired protection of deep networks from adversarial attacks
• Hierarchical interpretations for neural network predictions
• 3D ShapeNets: A Deep Representation for Volumetric Shapes
• Analyzing the Robustness of Nearest Neighbors to Adversarial Examples
• Towards Interpretable Deep Neural Networks by Leveraging Adversarial Examples
• A Fourier Perspective on Model Robustness in Computer Vision
• DeepCloak: Masking Deep Neural Network Models for Robustness Against Adversarial Samples
• Improving Adversarial Robustness via Channel-wise Activation Suppressing
• Explaining Deep Neural Networks with a Polynomial Time Algorithm for Shapley Values Approximation
• Explaining Explanations: Axiomatic Feature Interactions for Deep Networks
• Interpreting Adversarial Examples by Activation Promotion and Suppression
• Proper Network Interpretability Helps Adversarial Robustness in Classification
• Learning Global Pairwise Interactions with Bayesian Neural Networks
• A Unified Approach to Interpreting and Boosting Adversarial Transferability
• ML-LOO: Detecting Adversarial Examples with Feature Attribution
• Technical Note: Game-Theoretic Interactions of Different Orders
• A Game-Theoretic Taxonomy of Visual Concepts in DNNs
• Can We Faithfully Represent Masked States to Compute Shapley Values on a DNN?
• Building Interpretable Interaction Trees for Deep NLP Models
• A Hypothesis for the Aesthetic Appreciation in Neural Networks
• A Singular Value Perspective on Model Robustness