Robust Generalization Gap

The robust generalization gap in deep learning refers to the significant discrepancy between a model's performance on adversarially perturbed training data and its performance on unseen, similarly perturbed test data. Current research focuses on understanding this gap through analyses of model stability, architectural choices (like comparing CNNs and Vision Transformers), and the impact of sparsity techniques during adversarial training. Addressing this gap is crucial for deploying robust machine learning models in real-world applications where adversarial attacks are a concern, as it directly impacts the reliability and trustworthiness of these systems.

Papers

October 10, 2024

Understanding Adversarially Robust Generalization via Weight-Curvature Index
Yuelin Xu, Xiao Zhang
Adversarial Example Adversarial Robustness Adversarial Perturbation Robust Generalization Gaussian Curvature Robust Generalization Gap

January 6, 2024

Data-Dependent Stability Analysis of Adversarial Training
Yihan Wang, Shuang Liu, Xiao-Shan Gao
Adversarial Example Adversarial Training Generalization Bound Robust Generalization Stability Analysis Robust Generalization Gap

September 28, 2022

Exploring the Relationship between Architecture and Adversarially Robust Generalization
Aishan Liu, Shiyu Tang, Siyuan Liang, Ruihao Gong, Boxi Wu, Xianglong Liu, Dacheng Tao
Adversarial Attack Adversarial Example Adversarial Training Architecture Design Human Relationship Robust Generalization Robust Generalization Gap

May 27, 2022

Why Robust Generalization in Deep Learning is Difficult: Perspective of Expressive Power
Binghui Li, Jikai Jin, Han Zhong, John E. Hopcroft, Liwei Wang
Deep Learning Visual Perspective Robust Training Robust Generalization Expressive Power Robust Generalization Gap

February 20, 2022

Sparsity Winning Twice: Better Robust Generalization from More Efficient Training
Tianlong Chen, Zhenyu Zhang, Pengjun Wang, Santosh Balachandra, Haoyu Ma, Zehao Wang, Zhangyang Wang
Adversarial Training Sparsity Increase Robust Generalization Robust Generalization Gap

Robust Generalization Gap

Papers

Understanding Adversarially Robust Generalization via Weight-Curvature Index

Data-Dependent Stability Analysis of Adversarial Training

Exploring the Relationship between Architecture and Adversarially Robust Generalization

Why Robust Generalization in Deep Learning is Difficult: Perspective of Expressive Power

Sparsity Winning Twice: Better Robust Generalization from More Efficient Training