DAWN: A Dual-Decoder Attention Network  with Weighted Loss for Robust Minority-Class  Learning

Ishan Dwivedi; Sandhya Yadav; Ankit Baghel; Pawan Kumar Chaurasia

doi:10.36948/ijfmr.2026.v08i02.73751

DAWN: A Dual-Decoder Attention Network with Weighted Loss for Robust Minority-Class Learning

Author(s)	Mr. Ishan Dwivedi, Ms. Sandhya Yadav, Mr. Ankit Baghel, Dr. Pawan Kumar Chaurasia
Country	India
Abstract	Class imbalance remains a fundamental challenge in deep learning, where minority-class samples are often underrepresented and consequently overlooked during training. Existing approaches, including oversampling, re-weighted loss functions, and generic backbone architectures, provide only partial mitigation and frequently suffer from instability, overfitting, or inadequate representation of rare classes. To address these limitations, we propose DAWN (Dual-Decoder Attention with Weighted-Loss Network), a novel architecture explicitly designed for imbalance-aware learning. DAWN incorporates a dual-decoder framework, wherein a global decoder captures holistic semantic information, while a detail decoder focuses on fine-grained and minority-class features. A cross-attention alignment mechanism facilitates effective information exchange between the two decoders, preventing suppression of minority signals. Furthermore, a hybrid weighted loss function is introduced, integrating class-balanced optimization, overlap-aware objectives, and edge-aware constraints to achieve a balanced trade-off between minority sensitivity and overall accuracy. A targeted class-aware sampling strategy is also employed to mitigate rare-class forgetting during training. Extensive experiments conducted on five diverse datasets such as CIFAR-10-LT, CIFAR-100-LT, MNIST-imb, HAM10000 (medical imaging), and Credit Card Fraud(tabular anomaly detection) demonstrate the consistent superiority of DAWN over state-of-the-art baselines, achieving up to 10.2%improvement in macro-F1 score and 7.8% increase in minority-class recall, with less than 10% additional computational overhead. Statistical analysis confirms that these improvements are significant (p < 0.01) and robust across multiple runs. Qualitative evaluations further reveal sharper segmentation boundaries and interpretable decoder-specific attention maps. These findings establish DAWN as a powerful and generalizable framework for imbalance-aware deep learning, with strong applicability in domains where rare-category detection is critical, such as healthcare diagnostics, fraud detection, and cybersecurity.
Keywords	Class Imbalance, Long-Tailed Learning, Minority-Class Detection, Dual Decoder Networks, Attention Mechanisms, Hybrid Loss Functions, Cross-Attention, Medical Imaging, Fraud Detection, Deep Learning.
Field	Computer > Artificial Intelligence / Simulation / Virtual Reality
Published In	Volume 8, Issue 2, March-April 2026
Published On	2026-04-06
DOI	https://doi.org/10.36948/ijfmr.2026.v08i02.73751

View / Download PDF File

E-ISSN 2582-2160

doi

CrossRef DOI is assigned to each research paper published in our journal.

IJFMR DOI prefix is
10.36948/ijfmr

Downloads

Research Paper Format Copyright Permission Form and Undertaking Form Cover Page Vol 8 Isu 2 Cover Page Vol 8 Isu 1 Cover Page Vol 7 Isu 6

All research papers published on this website are licensed under Creative Commons Attribution-ShareAlike 4.0 International License, and all rights belong to their respective authors/researchers.

CC-BY-SA

About IJFMR Fees & Payment Current Issue Publication Archive	Submit Research Paper Track Submission Status Publication Guidelines Publication Ethics Peer Review & Plagiarism	Join as a Reviewer Editors & Reviewers Reviewer Referral Program Get Reviewer Membership Certi.	Website/Journal Policies Usage Policy Content Policies Privacy Policy

Contact Us		+91-9687-828-838	editor@ijfmr.com

International Journal For Multidisciplinary Research

A Widely Indexed Open Access Peer Reviewed Multidisciplinary Bi-monthly Scholarly International Journal

DAWN: A Dual-Decoder Attention Network with Weighted Loss for Robust Minority-Class Learning

Share this