Tiny transformers: multimodal digital biometric authentication model using tiny enhanced dual transformers with quality-gated fusion network

Biometric authentication remains a challenge due to inconsistencies arising from modality-specific noise, low-quality input samples, and overfitting when trained on small-scale datasets. Most existing models rely on shallow fusion strategies or task-specific architectures that fail to generalise across domains, mainly when modalities like iris and fingerprint differ in spatial structure and acquisition quality. To address these limitations, we propose a robust deep learning-based multimodal authentication pipeline framework that uses both iris and fingerprint modalities for personal digital authentication. Our architecture integrates Tiny Vision Transformer (Tiny ViT) backbones (Vision Transformer (ViT) and Swin) with a learned quality gated fusion technique, and angular margin-based metric learning employs ArcFace loss. This system utilises a comprehensive preprocessing pipeline for iris and fingerprint. For iris, we use the Hough transform and Contrast Limited Adaptive Histogram Equalization (CLAHE)-based contrast enhancement. For fingerprint, we use a Gabor filter with a morphological thinning process. To handle the low-quality images, we use Laplacian sharpness metrics, which are computed and incorporated into the gating network. This adaptively weights modality contribution during the feature fusion process. Two-layer Multi-Layer Perceptron (MLP) output weights are normalised with softmax, and score fusion is done. These novel tiny ViT and Swin Transformer models are evaluated using the cosine annealing scheduler with leave-one-out cross-validation. Our final gated fusion with ArcFace demonstrated strong performance, which achieves generalisation, with a 95% genuine acceptance rate at 1% false acceptance rate, an area under the curve (AUC) of 0.99, and 5% equal error rate (EER). This study demonstrates that adapting transformer-based backbones, using margin-based metric losses, and incorporating quality-gated fusion can yield interpretable and high-performing authentication systems, even under limited data scenarios.