Metric Learning for Adversarial Robustness

2019/09/03 by Chengzhi Mao, Mao, Chengzhi, Ziyuan Zhong +7 · 23 citations
Biochemistry, Genetics and Molecular Biology · Computer Science · Mathematics · #Adversarial Robustness in Machine Learning #Bacillus and Francisella bacterial research #cs.CR #cs.CV #cs.IR #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.1909.00900

arxiv created 2019/10/28 · arxiv updated 2019/10/29

Abstract

Deep networks are well-known to be fragile to adversarial attacks. We conduct an empirical analysis of deep representations under the state-of-the-art attack method called PGD, and find that the attack causes the internal representation to shift closer to the "false" class. Motivated by this observation, we propose to regularize the representation space under attack with metric learning to produce more robust classifiers. By carefully sampling examples for metric learning, our learned representation not only increases robustness, but also detects previously unseen adversarial samples. Quantitative experiments show improvement of robustness accuracy by up to 4% and detection efficiency by up to 6% according to Area Under Curve score over prior work. The code of our work is available at https://github.com/columbia/Metric_Learning_Adversarial_Robustness.

Cited by

• Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks
• Identifying Layers Susceptible to Adversarial Attacks
• Testing Machine Translation via Referential Transparency
• Recent Advances in Adversarial Training for Adversarial Robustness
• Semantics-Preserving Adversarial Training
• Bag of Tricks for Adversarial Training
• Improving Model Robustness with Latent Distribution Locally and Globally
• Defective Convolutional Networks
• Enhanced countering adversarial attacks via input denoising and feature restoring
• Feature Distillation With Guided Adversarial Contrastive Learning
• Online Adversarial Purification based on Self-Supervision
• Live Trojan Attacks on Deep Neural Networks
• RobustBench: a standardized adversarial robustness benchmark
• Model-Agnostic Meta-Attack: Towards Reliable Evaluation of Adversarial Robustness
• THAT: Two Head Adversarial Training for Improving Robustness at Scale
• Boosting Adversarial Training with Hypersphere Embedding
• Adversarial Robustness under Long-Tailed Distribution
• Diversity can be Transferred: Output Diversification for White- and Black-box Attacks
• Robustified Domain Adaptation
• Learnable Boundary Guided Adversarial Training
• Understanding and Achieving Efficient Robustness with Adversarial Supervised Contrastive Learning
• Structure-Preserving Progressive Low-rank Image Completion for Defending Adversarial Attacks
• Large-Scale Adversarial Training for Vision-and-Language Representation Learning

Related

• Improving Adversarial Robustness to Sensitivity and Invariance Attacks with Deep Metric Learning
• Enhancing Adversarial Robustness for Deep Metric Learning
• Towards Adversarially Robust Deep Metric Learning
• Provably Robust Metric Learning
• Improving Adversarial Robustness by Putting More Regularizations on Less Robust Samples