Adaptive Weight Assignment for Adversarial Training Based on Predicted Class Probabilities Across Different Attacks and Perturbation Sizes
Atsague, Modeste Tian, Jin Fakorede, Olukorede
Atsague, Modeste
Tian, Jin
Fakorede, Olukorede
Supervisor
Department
Machine Learning
Embargo End Date
Type
Conference proceeding
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Adversarial training (AT) improves model robustness by incorporating adversarial examples during training. Traditional methods, however, treat all examples equally, limiting their effectiveness. Recent studies show that adversarial examples vary in importance, and failing to account for this can weaken robustness. New approaches assign different weights to adversarial examples, improving defenses against specific attacks while maintaining natural accuracy. However, existing reweighting strategies often struggle against stronger attacks like CW and AA. Our analysis reveals that misclassified inputs may be assigned to different incorrect classes depending on the attack type and perturbation size, suggesting that more than one metric for weight assignment is required. To tackle this, we propose an Adaptive Weight Assignment (AWA) strategy that uses predicted class probabilities across multiple attack types and perturbation sizes. This method strengthens weaker adversarially trained models and significantly improves robustness against strong attacks like CW and AA, as confirmed by our extensive experiments.
Citation
M. Atsague, J. Tian, and O. Fakorede, “Adaptive Weight Assignment for Adversarial Training Based on Predicted Class Probabilities Across Different Attacks and Perturbation Sizes,” pp. 27–38, 2025, doi: 10.1007/978-981-96-8183-9_3
Source
Advances in Knowledge Discovery and Data Mining
Conference
29th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2025
Keywords
Subjects
Source
29th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2025
Publisher
Springer Nature