| Title |
Enhancing Adversarial Fairness via Data Pruning |
| Authors |
김성민(Seong Min Kim) ; 송병철(Byung Cheol Song) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.11.116 |
| Keywords |
Adversarial robustness; Adversarial fairness; Attack and defense; Computer vision |
| Abstract |
Recent studies have revealed that deep learning models are vulnerable to adversarial attacks. To ensure the safe deployment of deep learning in security-critical domains such as healthcare and autonomous driving, improving adversarial robustness is essential. Among various approaches, adversarial training has emerged as a fundamental method for enhancing robustness. However, adversarial training inherently suffers from adversarial fairness?a discrepancy in accuracy across different classes. While numerous studies have attempted to mitigate this issue by modifying the min-max framework, the relationship between adversarial unfairness and the underlying data has been largely overlooked. In this paper, we demonstrate the existence of data samples that exacerbate adversarial unfairness and show that adversarial fairness can be improved by pruning such harmful data from the training set. |