The Spatial Attention, illustrated in Fig. 2, functions similarly to the SE Attention; however, it is designed to focus on spatial features. Given an intermediate feature map F, spatial information is aggregated by applying average pooling and max pooling across the channel dimension, producing two spatial context descriptors, Favg∈R1×h×w and Fmax∈R1×h×w.(1)M(F)=σ(An×n([Favg;Fmax]))

These context descriptors are then combined to generate the spatial attention map M(F)(1) using the sigmoid function σ(x)=11+e−x. Here, An×n represents a n×n convolution kernel used to aggregate local spatial information, with smaller n capturing local details and larger n incorporating broader spatial context.

Instead of relying on closed-set classification, an embedding-based approach suited to the open-set nature of biometric recognition is adopted. As shown in Fig. 3, its design consists of a linear layer followed by batch normalization, producing compact feature embeddings. In related biometric tasks, embedding-based methods outperform classification-based ones: ArcFace achieved state-of-the-art performance, reporting 99.83% accuracy on LFW and 96.98% VR@FAR of 10−6 on MegaFace [17]; SphereFace improved verification from 98.71% using softmax to 99.42% on LFW [18]; and CosFace obtained an accuracy of 99.73% on LFW and 97.6% on YTF [19]. These methods train deep networks with margin-based losses, such as angular margin in SphereFace, cosine margin in CosFace, and additive angular margin in ArcFace, to project features into a hyperspherical space that enhances inter-class separability and intra-class compactness.(2)ℓ2(v)=v‖v‖2=[v1,v2,…,vn]v12+v22+⋯+vn2

Let v be an embedding vector of dimension n. ℓ2 normalization (2) is applied by dividing each element of v by its norm. After normalization, the resulting vectors satisfy unit length, ‖v‖2=1, giving all embeddings the same magnitude. This property ensures that the model focuses on the direction of the embeddings rather than their scale, since the direction encodes the identity-related information while the magnitude mostly reflects feature scale.

Triplet-based learning is employed to learn discriminative embeddings by enforcing a relative distance constraint among anchor, positive and negative samples. Specifically, the model is trained to minimize the Euclidean distance d between the anchor a and positive p while maximizing the distance between the anchor a and negative n at least a margin m, as formalized in the triplet margin loss Ltriplet (3). This encourages the embedding space to preserve identity-level similarity, ensuring that samples of the same class are closer together than samples of different classes. (3)Ltriplet=1N∑i=1Nmax(d(ai,pi)−d(ai,ni)+m,0)

To optimize the triplet margin loss, a margin scheduling strategy (5) is adopted, where the margin dynamically changes during training. The margin is initially fixed during a warm-hard phase and then gradually transitions to semi-hard phases using a cosine annealing schedule (4), facilitating effective integration with the hard negative mining strategy. Let m(t) denote the margin at epoch t, where Mwarm is the fixed warm-hard margin, and Mmedi, Mpost=[M0, M1] are the start and end margins for the first and second halves of the semi-hard phase, respectively. (4)CosAnneal=M0−12(M0−M1)(1−cos⁡(πt−TstartTend−Tstart)) (5)m(t)={Mwarm,1≤t≤TwarmCosAnneal(Mmedi),Twarm<t≤TmediCosAnneal(Mpost),Tmedi<t≤Tpost

To enhance convergence and embedding discriminability, hard negative mining strategy (6) is applied during training. In the initial warm-up stage, WarmHardNegativeMining is used, where semi-hard negatives that beyond the positive but within slightly loosened boundaries δleft to avoid overly difficult cases are selected. In the remaining stages of the training, SemiHardNegativeMining is used, where both δleft and δright are gradually tightened to focus on negatives closer to the anchor, thereby progressively increasing the difficulty of the sampled pairs. When no samples satisfy the warm-hard or semi-hard condition, the top-k hardest negatives are used as substitutes. This progressive strategy enables the model to gradually transition from easier to more challenging samples. (6)𝒩hard={Warm-hard:d(a,p)+loosen(δleft)<d(a,ni)<d(a,p)+δrightSemi-hard:d(a,p)+tighten(δleft)<d(a,ni)<d(a,p)+tighten(δright)Topk-hardest:d(a,ni)≤d(a,ni+1)≤⋯≤d(a,nk)

To strengthen the discriminative capability of the student model, MobileNetV3-L, knowledge distillation is employed using a pretrained teacher model, ResNet-50, trained on the same dataset, as shown in Fig. 4. The student model is optimized with a combined objective Ltotal (7), including a triplet loss Ltriplet (3) and a cosine similarity-based loss Ldistill (8) scaled by a factor δ for each embedding type x, and an overall factor α (9).(7)Ltotal=Ltriplet+α3∑x∈{a, p, n}δxLdistillx (8)Ldistillx=1N∑i=1N(1−cosine_similarity(six, tix))

Let α(t) denote the dynamic weighting factor at epoch t, with αsemi=[α0, α1] for controlling the contribution of teacher model during the semi-hard phase. During the initial warm-up phase, only Ltriplet is applied to allow the student model to first learn basic features without being influenced by the teacher model. Then, each Ldistillx is added to progressively align the student representations with the teacher. (9)α(t)={0,0<t≤TwarmCosAnneal(αsemi),Twarm<t≤Tsemi

A key step in recognition tasks is measuring the similarity between two normalized embedding vectors, fa,fb∈Rdim, using either Cosine similarity (10) or Euclidean distance (11). The similarity scores, sab and dab, are used to compute threshold-based metrics such as EER and VR@FAR, for verification, and ranking-based metrics such as CMC@K, for identification. (10)sab=fa⋅fb‖fa‖2‖fb‖2=∑k=1dimfa,kfb,k‖fa‖2‖fb‖2 (11)dab=‖fa−fb‖2=∑k=1dim(fa,k−fb,k)2

Verification metrics are used to evaluate the model’s ability to determine whether pairs of samples correspond to the same or different identities. Given a similarity score sab or distance dab, a threshold τ is applied to classify a pair as a match or non-match.

False Acceptance Rate, defined as FAR(τ)=NFA(τ)Nimp, measures the proportion of imposter pairs that are incorrectly accepted as matches, where NFA(τ) denotes the number of imposter pairs whose similarity exceeds the threshold τ (or distance is below τ), and Nimp is the total number of imposter pairs.

Identification metrics are used to evaluate the model’s ability to correctly identify a probe iris image from a gallery of known identities. In contrast to verification, which makes a binary decision, identification ranks all gallery images according to their similarity to the probe.

Cumulative Matching Characteristic, defined as CMC@k=Nprobe(k)Nprobe, measures the proportion that the correct identity of a probe appears is retrieved within the top-k ranked gallery results, where k denotes the rank threshold, Nprobe is the total number of probe images, and Nprobe(k) is the number of probes whose correct match appears in the top-k gallery images.

To evaluate the impact of the proposed optimizations, the baseline MobileNetV3-L (BASE) is compared with its optimized variants, and the state-of-the-art ResNet-50 (SOTA). Specifically, attention mechanism and knowledge distillation technique are incorporated into the baseline to create more robust variants. As observed, the individual variants show moderate improvements, whereas their combination (our proposal) yields a more substantial improvement and narrows the performance gap to the state-of-the-art. Table 3 presents a detailed comparison among the baseline, the optimized variants, and the state-of-the-art model.

The baseline achieves an EER of 1.699%, a VR@FAR of 97.678%, and a CMC@K of 93.073%, while remaining highly efficient at only 11.9 MB and 25.2 ms latency. In contrast, the state-of-the-art outperforms the baseline in accuracy, achieving an EER of 1.567%, a VR@FAR of 98.092%, and a CMC@K of 95.103%, at the expense of substantially larger size of 90.9 MB and higher latency of 188.1 ms. This comparison reflects the inherent trade-off between recognition accuracy and computational efficiency. To overcome this limitation, attention mechanism and knowledge distillation technique are applied, with the goal of pushing its accuracy closer to the state-of-the-art performance while retaining its compact-size and low-latency characteristics.

Compared to the baseline, both attention and distillation enhancements achieve consistent accuracy improvements with minimal computational cost. The attention-only variant reduces the EER to 1.484% and increases both the VR@FAR and CMC@K to 97.993% and 93.949%, respectively, as the attention mechanism helps the model capture more discriminative features. Similarly, the distilled-only variant achieves an EER of 1.535%, a VR@FAR of 97.951%, and a CMC@K of 94.466%, showing the benefit of knowledge transfer in enhancing feature discrimination. Taken together, these results confirm that each enhancement, regardless of the approach, independently strengthens the model’s discriminative ability.

Based on these individual enhancements, combining attention and distillation further improves overall performance. Compared to the baseline, the combined variant slightly increases model size by 0.1 MB and latency by 2 ms, while achieving superior performance across all metrics, with an EER of 1.409%, a VR@FAR of 98.184%, and a CMC@K of 94.785%. Moreover, it remains over 7× smaller and faster than the state-of-the-art model, achieving comparable recognition accuracy. Additionally, the combined model exhibits clear convergence during training, as shown in Fig. 5. These results highlight the complementary advantages of these two techniques, since the attention strengthens feature discrimination and the distillation fosters generalizable representations. Collectively, this combination provides an excellent trade-off between accuracy and efficiency, making the model highly practical for mobile iris recognition.

To further examine the practicality of the proposed models for mobile deployment, we assess their performance under post-training quantization. As summarized in Table 4, the results highlight that reduced numerical precision substantially impacts recognition accuracy and computational efficiency. As expected, quantization causes a noticeable degradation in recognition performance across all models.

In this work, post-training quantization is applied to the final models due to its simplicity, compatibility with mobile deployment, and no retraining requirement. Specifically, the weights are converted from float32 to float16, while other tensors remain in full precision due to the model’s sensitivity to numerical precision [20]. As float16 naturally preserves the dynamic range of the weights, converting them to half-precision reasonably reduces memory footprint and computational cost while minimizing precision loss.

The baseline model experiences a substantial accuracy drop after quantization, with an EER increasing from 1.699% to 7.587%, VR@FAR of 77.512% (a drop of about 20%), and CMC@K of 67.515% (a drop of about 27%), yet it retains a more compact size of 5.93 MB and lower latency of 12.7 ms. By comparison, the quantized state-of-the-art ResNet-50 remains more accurate but is significantly larger and slower, which is expected given the typical trade-off between accuracy and computational cost.

Meanwhile, the attention-only and distilled-only variants show slight overall improvements compared to the baseline, showing some resilience to quantization. Importantly, the combined variant achieves a balanced performance among all optimized models, with an EER of 7.173%, a VR@FAR of 79.726%, and a CMC@K of 70.143%, while maintaining a moderate size of 5.94 MB and reasonable latency of 13.2 ms. Overall, this indicates that integrating both enhancements effectively preserves discriminative features and provides a well-balanced trade-off between accuracy and efficiency, demonstrating its practicality for mobile deployment.

In general, the proposed optimizations enhance accuracy and maintain efficiency relative to the baseline; however, all models suffer a noticeable drop in recognition accuracy after quantization, with EER increasing, and VR@FAR and CMC@K decreasing across all variants. To address this, alternative strategies such as quantization-aware training or mixed precision could be considered to reduce the performance drop, though they involve additional training complexity and are left for future work. Consequently, this highlights the need for further development to better preserve accuracy and improve practicality for mobile deployment.