Deep iris recognition models are often trained on Cartesian grids, whereas iris texture follows a concentric structure with angular periodicity. This representational mismatch can weaken rotation robustness and limit pupil-to-limbus context modeling, while many pipelines still rely on accurate segmentation masks. We propose RadialFormer, an efficient mask-free iris recognition framework that performs representation learning directly in the polar domain. The pipeline first estimates pupil/iris parameters $(c_x,c_y,r_{\text{in}},r_{\text{out}})$ using a percentile radial-gradient operator with anatomical constraints, and then applies a crop-based polar transform to obtain a compact $64\times 512$ unwrapped iris map. To better match polar geometry, we introduce Learnable Polar Position Encoding (LPPE) with separable radial–angular embeddings, where Fourier terms in the angular branch enforce continuity at $\theta =0/2\pi$. We further propose Radial Stripe Window Attention (RSWA), which computes self-attention within full-height radial stripes and uses modular angular shifting to preserve circular consistency. Trained end-to-end with batch-hard triplet loss under $P\times K$ sampling, RadialFormer achieves 99.04% TPR@1%FPR with 0.48% EER on CASIA-V4-Lamp, and 93.63% TPR@1%FPR with 2.92% EER on CASIA-V4-Interval. Ablation and cross-dataset evaluations further validate the contributions of polar processing, LPPE, and RSWA and demonstrate robust generalization across acquisition conditions. Under the same input resolution, RadialFormer reduces computation by about $3.5\times$ compared with a standard transformer baseline while maintaining competitive recognition accuracy.

This paper introduces MobiIris, a lightweight deep network for mobile iris recognition that enhances attention and specifically addresses the balance between accuracy and efficiency on devices with limited resources. The proposed model is based on the large version of MobileNetV3 and adds more spatial attention blocks and an embedding-based head that was trained using margin-based triplet learning, enabling fine-grained modeling of iris textures in a compact representation. To further improve discriminability, we design a training pipeline that combines dynamic-margin triplet loss, a staged hard/semi-hard negative mining strategy, and feature-level knowledge distillation from a ResNet-50 teacher. Finally, we investigate the use of post-training float16 quantization to reduce memory footprint and latency for deployment on mobile hardware. Experiments on the challenging CASIA-IrisV4-Thousand dataset show that the full-precision MobiIris model requires only $12$ MB of storage and $27$ ms inference latency, while achieving an EER of $1.409\mathrm{\%}$, VR@FAR = $1\mathrm{\%}$ of $98.184\mathrm{\%}$, and CMC@ $1$ of $94.785\mathrm{\%}$, closely matching a ResNet-50 baseline that is more than $7\times$ larger and slower. Under post-training quantization, the model shrinks to $5.94$ MB with $13$ ms latency and maintains a competitive balance between accuracy and efficiency compared to other optimized variants. These results demonstrate that a coherent combination of lightweight architecture design, attention mechanisms, metric-learning objectives, hard negative mining, and knowledge distillation yields a practical iris recognition solution suitable for secure, real-time authentication on mobile and embedded platforms.