Interpreting Adversarial Examples with Attributes

Abstract

Deep computer vision systems being vulnerable to imperceptible and carefully crafted noise have raised questions regarding the robustness of their decisions. We take a step back and approach this problem from an orthogonal direction. We propose to enable black-box neural networks to justify their reasoning both for clean and for adversarial examples by leveraging attributes, i.e. visually discriminative properties of objects. We rank attributes based on their class relevance, i.e. how the classification decision changes when the input is visually slightly perturbed, as well as image relevance, i.e. how well the attributes can be localized on both clean and perturbed images. We present comprehensive experiments for attribute prediction, adversarial example generation, adversarially robust learning, and their qualitative and quantitative analysis using predicted attributes on three benchmark datasets.

References in corpus (14)

• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
• Towards Deep Neural Network Architectures Robust to Adversarial Examples
• Sanity Checks for Saliency Maps
• On Evaluating Adversarial Robustness
• Robustness May Be at Odds with Accuracy
• On Detecting Adversarial Perturbations
• Towards Interpretable Deep Neural Networks by Leveraging Adversarial Examples
• Interpretation of Neural Networks is Fragile
• Adversarial Images for Variational Autoencoders
• Robustness of classifiers to universal perturbations: a geometric perspective
• Adversarial Spheres
• Techniques for Interpretable Machine Learning
• Interpretable Deep Learning under Fire
• A Large-scale Attribute Dataset for Zero-shot Learning

Cited by in corpus (2)

• Unifying Model Explainability and Robustness via Machine-Checkable Concepts
• Deriving Explanation of Deep Visual Saliency Models