# A Data-Centric Approach for Improving Adversarial Training Through the Lens of Out-of-Distribution Detection

**Authors:** Mohammad Azizmalayeri, Arman Zarei, Alireza Isavand, Mohammad Taghi Manzuri, Mohammad Hossein Rohban  
**arXiv:** [2301.10454](https://arxiv.org/abs/2301.10454)  
**Date:** 2023-01-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.10454v1/

## Key points

- Researchers address the issue of adversarial perturbations in machine learning models
- Adversarial training is identified as an effective solution, where models are trained using adversarially perturbed samples
- Existing model-based methods can be equivalent to applying smaller perturbation or optimization weights to hard training examples
- Researchers propose a data-centric approach by detecting and removing hard samples directly from the training procedure
- Maximum softmax probability is used for out-of-distribution detection to identify hard samples
- Experiments conducted on SVHN and CIFAR-10 datasets show significant improvement in adversarial training by removing hard samples
- The proposed method enhances model robustness against adversarial perturbations.

## Summary

In this study, the researchers address the issue of adversarial perturbations in machine learning models. While these models have achieved super-human performance in various real-world applications, they are still vulnerable to imperceptible adversarial perturbations. Adversarial training has been identified as an effective solution to this problem, where the model is trained using adversarially perturbed samples instead of original ones. To improve adversarial training, several methods have been developed over recent years, such as data augmentation or modifying training attacks. However, in this work, the researchers propose a new data-centric approach to tackle this problem. They first demonstrate that existing model-based methods can be equivalent to applying smaller perturbation or optimization weights to hard training examples. Building on this finding, the researchers suggest detecting and removing these hard samples directly from the training procedure instead of using complicated algorithms to mitigate their effects. For detection, they utilize maximum softmax probability as an effective method for out-of-distribution detection since hard samples can be considered as out-of-distribution samples for the entire data distribution. The effectiveness of this approach is demonstrated through experiments conducted on SVHN and CIFAR-10 datasets. The results show that by removing hard samples from the training process, adversarial training can be significantly improved without adding excessive computational cost. Overall, this study presents a novel perspective on improving adversarial training by focusing on data-centric approaches and effectively detecting and removing hard samples during the training procedure. The proposed method shows promising results in enhancing model robustness against adversarial perturbations.