Exploring Neural Network Robustness: Evaluating Reaction to Image Data Perturbations
IEEE Access · 2026 · 14 · 81768-81782
DOI: 10.1109/access.2026.3695335 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Damian Raczkowski und 1 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
While Deep Neural Networks have achieved superhuman performance in image classification, their fragility against minor data corruptions remains a critical barrier to real-world deployment in safety-critical fields such as healthcare and autonomous driving. This study moves beyond standard accuracy benchmarks to conduct a systematic “stress test” on five distinct architectural paradigms: VGG16, ResNet152, ConvNeXt, EfficientNet-V2-M, and Vision Transformers. In contrast to conventional corruption benchmarks that assess only output accuracy, the proposed approach quantifies representational drift using Mean Squared Error between clean and perturbed feature vectors and analyses class-level failure convergence. Experimental results reveal substantial architecture-dependent robustness differences. Under Gaussian noise, VGG16 collapses after only 10% pixel corruption, whereas EfficientNet-V2-M maintains partial recognition up to 50% corruption. Under severe Gaussian blur, EfficientNet retains over 63% top-1 accuracy, while other architectures achieve at most approximately 39%. The analysis further identifies attractor-class behaviour, where models converge to specific incorrect predictions (e.g., “scarf” or “kite”) under strong perturbations. Moreover, a representational–decision discrepancy is observed in Vision Transformers, which preserve comparatively high classification accuracy despite significantly larger feature-space instability (MSE approximately 0.2 higher than ResNet, ConvNeXt, and EfficientNet). This work provides a granular map of architectural vulnerabilities, offering crucial insights for designing models that can withstand the imperfections of real-world applications. Project’s code is available on GitHub: https://github.com/DocentSzachista/skrypt-badania-magisterka
Zitieren
APA
Raczkowski, D., & Szandała, T. (2026). Exploring Neural Network Robustness: Evaluating Reaction to Image Data Perturbations. IEEE Access, 14, 81768-81782. https://doi.org/10.1109/access.2026.3695335
Vancouver
Raczkowski D, Szandała T. Exploring Neural Network Robustness: Evaluating Reaction to Image Data Perturbations. IEEE Access. 2026;14:81768-81782. doi:10.1109/access.2026.3695335
BibTeX
@article{Raczkowski2026,
author = {Raczkowski, Damian and Szandała, Tomasz},
title = {Exploring Neural Network Robustness: Evaluating Reaction to Image Data Perturbations},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {81768--81782},
doi = {10.1109/access.2026.3695335}
}
EUVIMED