Abstract—This study investigates the role of explainability in improving adversarial robustness of Convolutional Neural Network (CNN)-based face recognition systems. While prior works treat robustness and interpretability as separate objectives, we conduct a systematic analysis of how explanation-driven preprocessing can influence adversarial resilience. Specifically, we combine adversarial training with Local Interpretable Model Agnostic Explanation (LIME)-guided feature perturbation to identify and suppress vulnerable regions. Unlike prior approaches that primarily use Explainable Artificial Intelligence (XAI) for post-hoc interpretation, this work quantitatively evaluates the relationship between explanation consistency and robustness under adversarial perturbations. Experimental results on CASIA-WebFace and Labeled Faces in the Wild (LFW) show that the proposed framework improves robustness under low-strength attacks (ε = 0.02), increasing Area Under Curve (AUC) from 62.21% to 75.76%. However, performance degrades under stronger perturbations, indicating limitations of explanation-guided preprocessing. These findings provide empirical evidence that explainability can contribute to robustness, but its effectiveness is highly dependent on attack strength and preprocessing design. This work highlights both the potential and limitations of integrating XAI into adversarial defence.
 
Keywords—adversarial attack, adversarial training, face recognition, explainable artificial intelligence, Local Interpretable Model Agnostic Explanation (LIME)


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).