Proceedings of the International Conference on Intelligent Data Analysis and Applications (IDAA 2025)

<Previous Article In Volume
Next Article In Volume>

Dynamic Region-Aware Gradient Suppression (DRAGS): Enhancing Vision Model Robustness by Suppressing Noisy Feature Regions During Training

Authors
Md Muntaqim Meherab1, *, Nuruzzaman Faruqui2, Faria Nishat Khan3, Tanvirul Islam1, Syed Asif Johan4, Md. Maruf Billah5, Kazi Shakkhar Rahman6, Z. N. M. Zarif Mahmud1, Tauhidul As Sami7
1Dept. of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh
2Dept. of Software Engineering, Daffodil International University, Dhaka, Bangladesh
3PhD in Data Science and Engineering, South Dakota School of Mines and Technology, Rapid City, SD, USA
4Dept. of Computer Science and Engineering, Rajshahi University of Engineering and Technology, Rajshahi, Bangladesh
5Dept. of Electrical and Computer Engineering, North South University, Dhaka, Bangladesh
6Dept. of Computer Science and Engineering, Islamic University of Technology, Gazipur, Bangladesh
7Dept. of Biomedical Engineering, Bangladesh University of Engineering and Technology, Dhaka, Bangladesh
*Corresponding author. Email: meherab2305101354@diu.edu.bd
Corresponding Author
Md Muntaqim Meherab
Available Online 8 June 2026.
DOI
10.2991/978-94-6239-664-7_47How to use a DOI?
Keywords
Robustness; regularization; gradient gating; dynamic masking; corruptions; CIFAR-10; ResNet-18; computer vision
Abstract

Deep vision models can perform very well on clean test sets but still break down when inputs are corrupted by noise, blur, or occlusion. We introduce Dynamic Region-Aware Gradient Suppression (DRAGS), a lightweight training-time mechanism that suppresses gradients from spatial regions detected as noisy or spurious for the current mini-batch. DRAGS computes a saliency score at each location in an intermediate feature map, keeps only the top τ fraction, and detaches the gradients from the remaining locations. This dynamic, region-level gating differs from classical regularizers that act uniformly across space or channels. On CIFAR-10 with a ResNet-18 backbone, DRAGS yields a consistent improvement in clean accuracy over a strong baseline (+1.39 points; 64.15% ± 2.30 vs. 62.76% ± 1.05 across three seeds) and maintains robustness under common corruptions, with slight gains on Gaussian noise and occlusion and a measured trade-off under motion blur. A short ablation over τ and layer placement shows that moderate suppression on early layers is the most reliable setting. Heatmaps confirm that DRAGS downweights visually noisy background zones while preserving informative object regions. Overall, DRAGS is simple to implement and computeefficient relative to adversarial training, making it a practical option for robustness-minded training at scale.

Copyright
© 2026 The Author(s)
Open Access
Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

<Previous Article In Volume
Next Article In Volume>
Volume Title
Proceedings of the International Conference on Intelligent Data Analysis and Applications (IDAA 2025)
Series
Advances in Intelligent Systems Research
Publication Date
8 June 2026
ISBN
978-94-6239-664-7
ISSN
1951-6851
DOI
10.2991/978-94-6239-664-7_47How to use a DOI?
Copyright
© 2026 The Author(s)
Open Access
Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Cite this article

risenwbib
TY  - CONF
AU  - Md Muntaqim Meherab
AU  - Nuruzzaman Faruqui
AU  - Faria Nishat Khan
AU  - Tanvirul Islam
AU  - Syed Asif Johan
AU  - Md. Maruf Billah
AU  - Kazi Shakkhar Rahman
AU  - Z. N. M. Zarif Mahmud
AU  - Tauhidul As Sami
PY  - 2026
DA  - 2026/06/08
TI  - Dynamic Region-Aware Gradient Suppression (DRAGS): Enhancing Vision Model Robustness by Suppressing Noisy Feature Regions During Training
BT  - Proceedings of the International Conference on Intelligent Data Analysis and Applications (IDAA 2025)
PB  - Atlantis Press
SP  - 679
EP  - 687
SN  - 1951-6851
UR  - https://doi.org/10.2991/978-94-6239-664-7_47
DO  - 10.2991/978-94-6239-664-7_47
ID  - Meherab2026
ER  -