2021/12/01 by Hadi M. Dolatabadi, Sarah Erfani, Dolatabadi, Hadi M. +3 · 1 citation
Computer Science · Social Sciences · #Advanced Neural Network Applications #Adversarial Robustness in Machine Learning #Anomaly Detection Techniques and Applications #Computer Vision and Pattern Recognition (cs.CV) #Ethics and Social Impacts of AI #FOS: Computer and information sciences #Machine Learning (cs.LG) #Stochastic Gradient Optimization Techniques
paper · pdf · doi:10.48550/arxiv.2112.00378
openalex publication_date 2021/12/01 · openalex created_date 2022/07/22 · openalex updated_date 2026/07/28
Neural networks are vulnerable to adversarial attacks: adding well-crafted,\nimperceptible perturbations to their input can modify their output. Adversarial\ntraining is one of the most effective approaches in training robust models\nagainst such attacks. However, it is much slower than vanilla training of\nneural networks since it needs to construct adversarial examples for the entire\ntraining data at every iteration, hampering its effectiveness. Recently, Fast\nAdversarial Training (FAT) was proposed that can obtain robust models\nefficiently. However, the reasons behind its success are not fully understood,\nand more importantly, it can only train robust models for \ℓ_\∞-bounded\nattacks as it uses FGSM during training. In this paper, by leveraging the\ntheory of coreset selection, we show how selecting a small subset of training\ndata provides a general, more principled approach toward reducing the time\ncomplexity of robust training. Unlike existing methods, our approach can be\nadapted to a wide variety of training objectives, including TRADES,\n\ℓp-PGD, and Perceptual Adversarial Training (PAT). Our experimental\nresults indicate that our approach speeds up adversarial training by 2-3 times\nwhile experiencing a slight reduction in the clean and robust accuracy.\n