A hybrid approach exploiting machine learning models and SIFT-based matching for speed-optimized comparisons in iris authentication systems

Iris represents one of the most reliable biometric modalities due to its high distinctiveness and long-term stability. Classical iris-based authentication algorithms compare the user's iris with all the irises stored in the system, leading to high computational costs and scalability issues as the database grows. In this paper, we present a hybrid multi-algorithm system that integrates traditional image-matching techniques with modern machine learning approaches to enhance authentication accuracy and robustness. Our approach significantly reduces the number of comparisons, improving efficiency without compromising accuracy. The proposed hierarchical framework employs three classifiers – Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Feed-Forward-Neural-Network (FFNN) – and iris-codes extracted via Daugman's method. These classifiers effectively shrink the candidate pool, which is subsequently verified using the Scale-Invariant Feature Transform (SIFT) for precise feature matching. Moreover, by leveraging only non-deep techniques, our system is lightweight and well-suited for deployment on resource-constrained devices, such as mobile, wearable, and IoT devices. The integration of machine learning with SIFT enhances the overall performance and security of the biometric process, offering significant improvements in processing time, scalability, and deployability in real-world and resource-limited environments. To demonstrate the generalizability of the proposed approach, the system was rigorously evaluated across four heterogeneous datasets: CASIA-Iris v1, CASIA-Iris-Thousand v4, UBIRIS v1, and ND-Iris-0405. These benchmarks encompass a broad spectrum of image qualities, sample sizes, and environmental conditions. This solution achieved accuracies of 98.6%, 92.03%, 95.97%, and 97.3% for the CASIA-Iris v1, UBIRIS v1, ND-Iris 0405, and CASIA-Iris-Thousand v4 databases, respectively.